An intelligent operation and maintenance method and system for a clean room based on artificial intelligence

Abstract

The present application relates to the field of artificial intelligence and clean room intelligent operation and maintenance technology, in particular to a clean room intelligent operation and maintenance method and system based on artificial intelligence, comprising: collecting business data sources, Internet of Things data sources and spatial topology configuration data; based on production scheduling, personnel authority and standard operating procedures, equipment account and maintenance records and spatial configuration data, an ideal benchmark space-time tensor organized according to preset time granularity and spatial partition is constructed; abnormal event operators are used for parameterized disturbance to generate an abnormal simulation state set; the real collected data and the abnormal simulation state are respectively differentiated with the ideal benchmark to obtain a real residual error matrix and a theoretical residual error matrix set; a diagnosis result is generated based on time sequence and spatial similarity, and an operation and maintenance work order or an observation record to be observed is generated according to conditions; the execution and review results are collected, the disturbance parameters and the dispatch threshold are updated, and the accurate identification and disposal of personnel violations, equipment degradation and logistics retention are realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and cleanroom intelligent operation and maintenance technology, specifically to an artificial intelligence-based cleanroom intelligent operation and maintenance method and system. Background Technology

[0002] In the existing cleanroom operation and maintenance process, multiple types of data are usually integrated, such as production scheduling, personnel management, equipment ledger, access control records and environmental monitoring, to monitor personnel entry and exit, equipment operating status and environmental parameter changes in the clean area, so as to help determine whether the workshop operation deviates from the expected requirements.

[0003] However, most existing technologies rely on single-point threshold alarms, historical average comparisons, or isolated device alarms to identify anomalies. This makes it difficult to combine specific production batches, regional connectivity, personnel standard operating procedures, and equipment service areas to distinguish between reasonable changes caused by production scheduling changes, normal logistics transfers, and short-term environmental fluctuations, and non-standard deviations caused by personnel violations, equipment degradation, or spatial obstruction. This results in inaccurate identification of the root causes of anomalies, insufficient targeting of work orders, and poor closed-loop operation and maintenance. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent operation and maintenance method and system for cleanrooms based on artificial intelligence. Specifically, the technical solution of this invention is as follows:

[0005] An AI-based intelligent operation and maintenance method for cleanrooms includes: collecting business data sources, IoT data sources, and spatial topology configuration data, wherein the spatial topology configuration data includes at least area division, area connectivity, sensor point-to-area mapping, and equipment-to-area mapping; constructing an ideal reference spatiotemporal tensor organized according to preset time granularity and spatial partitions based on production scheduling data, personnel access and standard operating procedure data, equipment ledger and maintenance record data, and spatial topology configuration data; parametrically perturbing the ideal reference spatiotemporal tensor using anomaly event operators to generate an abnormal simulation state set; and performing a first difference between equipment power consumption data, access control trigger data, and environmental monitoring data and the spatial slices of the corresponding spatial partitions in the ideal reference spatiotemporal tensor. The actual residual matrix is ​​obtained, and the second difference is performed between each abnormal simulation state and the ideal reference spatiotemporal tensor to obtain the theoretical residual matrix set. Based on the temporal and spatial similarity between the actual residual matrix and the theoretical residual matrix set, a diagnostic result is generated, which includes the anomaly type and a comprehensive matching score. When the diagnostic result meets the preset dispatch conditions, an operation and maintenance work order is generated and a disposal instruction is sent; otherwise, an observation record is generated. The preset dispatch conditions include at least a comprehensive matching score threshold or an anomaly duration threshold. The execution result data and review result data of the operation and maintenance work order are collected, a judgment accuracy index is generated, and the perturbation parameters of the abnormal event operator and the comprehensive matching score threshold or anomaly duration threshold in the preset dispatch conditions are updated based on the judgment accuracy index.

[0006] Optionally, an ideal reference spatiotemporal tensor is constructed, including: parsing production scheduling data to obtain task time sequences; parsing personnel permissions and standard operating procedure data based on task time sequences to obtain standard behavior sequences; generating theoretical reference power consumption distribution, theoretical pressure difference distribution, and theoretical environmental load distribution based on task time sequences, standard behavior sequences, equipment ledgers and maintenance records, and spatial topology configuration data; and performing time alignment and spatial mapping on the theoretical reference power consumption distribution, theoretical pressure difference distribution, and theoretical environmental load distribution to obtain the ideal reference spatiotemporal tensor.

[0007] Optionally, the abnormal event operator includes time offset parameters, amplitude offset parameters, duration parameters, and spatial influence range parameters. The abnormal event operator is used to parameterize and perturb the ideal reference spatiotemporal tensor to generate a set of abnormal simulation states. This includes: calling an expert rule base to determine personnel violation event operators, equipment degradation event operators, and logistics congestion event operators; using the personnel violation event operator to perform personnel behavior perturbations to obtain personnel abnormal simulation states; using the equipment degradation event operator to perform equipment performance perturbations to obtain equipment abnormal simulation states; using the logistics congestion event operator to perform spatial obstruction perturbations to obtain flow field abnormal simulation states; and summarizing the personnel abnormal simulation states, equipment abnormal simulation states, and flow field abnormal simulation states to obtain a set of abnormal simulation states.

[0008] Optionally, the actual residual matrix is ​​a deviation matrix with time slices as rows and preset monitoring feature dimensions as columns. The preset monitoring feature dimensions include at least one of the following: area power consumption deviation, access control trigger count deviation, door opening duration deviation, particle concentration deviation, pressure difference deviation, and environmental recovery time deviation.

[0009] Each theoretical residual matrix in the theoretical residual matrix set has the same dimension as the actual residual matrix; the residuals in the first and second differences are absolute differences, relative deviations, or normalized differences; before generating diagnostic results, the following steps are also included: normalizing the actual residual matrix and the theoretical residual matrix set and splicing them according to time windows to construct a unified feature space.

[0010] Optionally, a diagnostic result is generated based on the temporal and spatial similarity between the actual residual matrix and the theoretical residual matrix set. This includes: calculating the temporal similarity between the actual residual matrix and each theoretical residual matrix using dynamic time warping; calculating the spatial similarity between the actual residual matrix and each theoretical residual matrix using cosine similarity; weighting and fusing the temporal and spatial similarities according to preset weights to obtain a comprehensive matching score; comparing the comprehensive matching score with preset high and low thresholds, where the preset high threshold is greater than the preset low threshold; outputting the corresponding abnormality type when the comprehensive matching score is greater than or equal to the preset high threshold; outputting a normal result when the comprehensive matching score is less than or equal to the preset low threshold; and outputting a review result when the comprehensive matching score is greater than the preset low threshold and less than the preset high threshold.

[0011] Optionally, the preset dispatch conditions include at least one or more of the following: anomaly type, comprehensive matching score threshold, anomaly duration threshold, and impact level threshold; generating maintenance work orders and sending disposal instructions, including: when the anomaly type is personnel violation, generating a retraining work order and sending it to the corresponding personnel terminal; when the anomaly type is equipment degradation, generating a maintenance work order and sending it to the equipment management terminal; when the anomaly type is spatial obstruction, generating a dispatch work order and sending it to the logistics management terminal; and storing the retraining work order, maintenance work order, or dispatch work order in association with the diagnostic results.

[0012] Optionally, the perturbation parameters of the abnormal event operator and the matching score threshold or duration threshold in the preset distribution conditions are updated based on the decision accuracy index, including: when the decision accuracy index is lower than the preset accuracy threshold, the time offset parameter, amplitude offset parameter, duration parameter or spatial influence range parameter of the abnormal event operator are adjusted, and the matching score threshold or duration threshold in the preset distribution conditions is adjusted; when the decision accuracy index is higher than or equal to the preset accuracy threshold, the current parameters and thresholds are retained.

[0013] An AI-based intelligent operation and maintenance system for cleanrooms includes a processor and a memory. The memory is coupled to the processor and is used to store computer program code, which includes computer instructions. The processor invokes the computer instructions to cause the system to perform the method as described in the embodiments.

[0014] Optionally, the system includes: a data acquisition module for acquiring business data sources, IoT data sources, and spatial topology configuration data; a benchmark reconstruction module for constructing an ideal benchmark spatiotemporal tensor; an anomaly perturbation module for generating a set of anomaly simulation states; a difference module for generating a set of real residual matrices and theoretical residual matrices; a decision module for generating diagnostic results; a work order dispatch module for generating maintenance work orders and sending disposal instructions; a feedback acquisition module for acquiring execution result data and review result data; and an evaluation and update module for generating decision accuracy indicators and updating the perturbation parameters of the anomaly event operator and the matching score threshold or duration threshold in the preset dispatch conditions.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. This invention collects business data sources, IoT data sources, and spatial topology configuration data, and constructs an ideal benchmark spatiotemporal tensor based on production scheduling data, personnel permissions and standard operating procedure data, and equipment ledgers and maintenance records. This method changes the existing technology's reliance on historical averages or single-point thresholds for comparison, and closely integrates the temporal sequence of specific production tasks with spatial connectivity, thereby accurately distinguishing between reasonable fluctuations caused by production scheduling changes or normal logistics transfers and non-standard deviations caused by real anomalies, solving the problem of inaccurate identification of anomaly root causes.

[0017] 2. This invention utilizes anomaly event operators to parametrically perturb the ideal reference spatiotemporal tensor, generating a set of anomalous simulation states, and further differencing to obtain a set of theoretical residual matrices. The system calls upon an expert rule base to specifically execute perturbations of personnel behavior, equipment performance, and spatial obstruction, constructing comparable theoretical damaged states. This enables the system to move beyond simple alarms lacking anomaly tracing capabilities and instead provide interpretable and refined qualitative analysis of complex non-standard deviations such as personnel violations, equipment degradation, and logistics delays.

[0018] 3. This invention compares the real residual matrix generated by fusing multi-source real data with the theoretical residual matrix set in the feature space. It uses dynamic time warping to calculate the temporal similarity of the temporal evolution beat and cosine similarity to calculate the cosine similarity of the spatial distribution. The weighted fusion yields a comprehensive matching score. This spatiotemporal dual-dimensional alignment calculation avoids misjudgment caused by the similarity of a single environmental parameter or isolated device indicators, making the diagnostic results in complex operation and maintenance scenarios more accurate and reliable.

[0019] 4. This invention generates a targeted maintenance work order and issues a disposal instruction when the diagnostic results meet the preset dispatch conditions, solving the problems of insufficient targeting and poor closed-loop nature of work order dispatch in traditional maintenance. At the same time, the system further collects the execution result data and review result data of the maintenance work order to generate a judgment accuracy index, and dynamically adjusts the disturbance parameters of the abnormal event operator and the various thresholds of the preset dispatch conditions based on the index. This enables the system to continuously self-correct and optimize based on long-term field feedback, and continuously conform to the actual operating rules of a specific cleanroom. Attached Figure Description

[0020] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0021] Figure 1 This is a flowchart of the method of the present invention;

[0022] Figure 2 This is a structural diagram of the system of the present invention. Detailed Implementation

[0023] 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.

[0024] Example 1:

[0025] Please see Figure 1An AI-based intelligent operation and maintenance method for cleanrooms includes: collecting business data sources, IoT data sources, and spatial topology configuration data, wherein the spatial topology configuration data includes at least area division, area connectivity, sensor point-to-area mapping, and equipment-to-area mapping; constructing an ideal reference spatiotemporal tensor organized according to preset time granularity and spatial partitions based on production scheduling data, personnel access and standard operating procedure data, equipment ledger and maintenance record data, and spatial topology configuration data; parametrically perturbing the ideal reference spatiotemporal tensor using anomaly event operators to generate an abnormal simulation state set; and performing a first difference between equipment power consumption data, access control trigger data, and environmental monitoring data and the spatial slices of the corresponding spatial partitions in the ideal reference spatiotemporal tensor. The actual residual matrix is ​​obtained, and the second difference between each abnormal simulation state and the ideal reference spatiotemporal tensor is performed to obtain the theoretical residual matrix set. Based on the temporal and spatial similarity between the actual residual matrix and the theoretical residual matrix set, a diagnostic result is generated, which includes the abnormality type and the comprehensive matching score. When the diagnostic result meets the preset dispatch conditions, an operation and maintenance work order is generated and a disposal instruction is sent; otherwise, an observation record is generated. The preset dispatch conditions include at least a comprehensive matching score threshold or an abnormality duration threshold. The execution result data and review result data of the operation and maintenance work order are collected, a judgment accuracy index is generated, and the perturbation parameters of the abnormal event operator and the comprehensive matching score threshold or abnormality duration threshold in the preset dispatch conditions are updated according to the judgment accuracy index.

[0026] This embodiment provides an intelligent operation and maintenance mechanism for cleanrooms based on artificial intelligence. Specifically, taking the injection filling cleanroom of a sterile pharmaceutical manufacturing company as a continuous main scenario, the workshop includes at least a personnel changing area, a material buffer zone, a filling core area, a capping area, and a finished product temporary storage area. There are clear access control connections, pressure gradient requirements, and equipment layout relationships between each area. The system is deployed in the workshop's operation and maintenance management server and is connected to the manufacturing execution system, enterprise resource planning system, access control system, building management system, and environmental monitoring network. It is used to distinguish between reasonable fluctuations caused by production scheduling and non-standard deviations caused by personnel violations, equipment degradation, or logistics delays during the execution of the same production batch.

[0027] Specifically, three types of data are collected: Business data sources may include the planned start time of a batch of vials filling tasks, the switchover time of each process, the list of personnel allowed to enter the core area, job operation steps, the equipment ledger of the fan / filter unit, and the last maintenance time; IoT data sources may include the power consumption of the fan / filter unit, access control opening and closing records, particle concentration, temperature and humidity, and pressure difference between adjacent areas; Spatial topology configuration data is used to describe the actual spatial organization of the workshop. For example, areas can be labeled as Z1 changing area, Z2 material buffer zone, Z3 filling core area, and Z4 capping area; if there is a transfer door between Z2 and Z3, the topology records the connectivity relationship between Z2 and Z3; if particle sensor P3 is installed in the filling core area, the mapping between P3 and Z3 is recorded; if fan / filter unit-7 serves the laminar flow unit above Z3, the mapping between fan / filter unit-7 and Z3 is recorded; through this topology, the system can understand where a door is opened, which adjacent area an anomaly will spread to, and which equipment will affect which area.

[0028] Based on this, the system constructs an ideal reference spatiotemporal tensor. Here, "ideal" does not refer to the historical average state, but rather to the baseline business scenario corresponding to the current batch of tasks under the conditions of fully complying with standard operating procedures, equipment being in good condition, and logistics flowing on time. Time can be sliced ​​in 5-minute segments, and space can be divided by region. For example, within time slices T1 to T4, if the plan shows that the material buffer zone completes a standard transfer to the filling core area, the system will provide the following corresponding values ​​in T2 and T3: access control opening time is less than or equal to the first preset time threshold, local environmental load increases but the pressure difference still maintains the design gradient, and the power consumption of the fan filter unit rises within an acceptable range and is greater than or equal to the preset power consumption baseline value. The baseline formed in this way is not an abstract data stacking, but rather the target power consumption and target environmental performance of the expected baseline operating state in this batch, this region, and this moment.

[0029] The system invokes anomaly event operators to perturb the aforementioned ideal baseline, generating a set of abnormal simulation states. These event operators can be understood as parameterized expressions of common anomalies in a real workshop. Personnel violations can manifest as personnel entering the core area without completing secondary dust removal as required after changing clothes, causing an increase in particle load in the core area, which then propagates to adjacent areas within a preset time window after the door opens. Equipment degradation can manifest as increased power consumption in the fan filter unit after dust accumulation, in order to maintain the designed airflow, but the local pressure difference recovery time exceeds the preset recovery time threshold. Logistics congestion can manifest as a transfer vehicle stopping at the junction of the passage and buffer zone, causing local flow obstruction and prolonging the duration of pressure difference fluctuations in the doorway area. The system superimposes these anomalies onto the ideal baseline one by one, forming multiple sets of theoretically damaged states.

[0030] Furthermore, the system compares the actual collected equipment power consumption data, access control trigger data, and environmental monitoring data with the ideal benchmark to obtain the actual residual matrix; simultaneously, it compares each group of theoretical damage states with the ideal benchmark to obtain the corresponding theoretical residual matrix. Here, the residual can be understood as the magnitude and direction of the deviation of the actual state or simulation state from the ideal business cycle. For example, if the plan is to allow doors Z2 to Z3 to be opened for 8 seconds at time T2, but the actual record is 20 seconds, and the duration of the Z3 particle concentration rise exceeds the preset benchmark threshold, and the power consumption of the fan filter unit -7 is also continuously higher than the preset power consumption benchmark value from T2 to T4, then the actual residual will simultaneously reflect the coordinated deviation of access control, environment, and energy consumption. Correspondingly, the theoretical residual obtained from the simulation of personnel violation events will also present a similar graph of particle rise after the door is opened, accompanied by short-term power consumption compensation; while the simulation of equipment degradation events is more likely to show long-term high power consumption, slow differential pressure recovery, but insignificant access control anomalies.

[0031] Based on this, the system calculates the temporal and spatial similarity between the actual residuals and each theoretical residual, and outputs the diagnostic results. The temporal similarity reflects whether the order in which the deviations occur is consistent, such as whether they all show the pattern of first opening the door, then raising the particles, and then compensating for power consumption. The spatial similarity reflects whether the distribution of the deviations in each region is consistent, such as whether the anomalies are mainly concentrated at the entrance of the filling core area rather than the entire corridor. If a theoretical anomaly is similar to the actual deviation in both temporal evolution and regional distribution, the system will give the corresponding anomaly type and comprehensive matching score.

[0032] When the diagnostic results meet the preset dispatch conditions, the system automatically generates an operation and maintenance work order and sends a disposal instruction; if not, it retains the observation record; for example, for deviations with an opening duration of less than the first preset duration threshold and a limited impact range, the system can only record them as observations to prevent interference with normal production rhythm; for deviations with a long duration and affecting the core area, a disposal task is automatically generated; after disposal, the system collects the work order execution results and review results, such as whether power consumption has recovered after equipment maintenance, whether similar anomalies still occur in subsequent shifts after personnel retraining, and whether pressure fluctuations have disappeared after logistics adjustments, and forms a judgment accuracy index; if a certain type of anomaly is frequently misjudged, the disturbance parameters of the corresponding event operator, as well as the matching score threshold or duration threshold in the work order dispatch threshold, are updated so that the system gradually conforms to the actual operating rules of the workshop;

[0033] As a supplementary processing mechanism, if a sensor fails to connect, access control logs are missing, or production scheduling changes are not synchronized during a certain time slice, the system can first mark the relevant time slice as a low-confidence segment, suspend automatic order dispatch for that time slice, and only output records to be observed or review suggestions; if the spatial topology configuration changes, such as temporarily closing a logistics channel, the system will not use the old topology to infer the spread of anomalies before the new topology is confirmed, so as to avoid misjudging the impact of the renovation and construction as an operational anomaly; if multiple theoretical anomalies have high similarity at the same time, the system can output the review results instead of forcibly classifying them into a single anomaly, so as to leave room for manual confirmation.

[0034] For example, in the middle of a batch of freeze-dried powder injection filling tasks, the plan requires the material to be transferred from the Z2 material buffer zone to the Z3 filling core zone. The access control standard opening time is less than the preset duration threshold. The particle concentration in the core zone is allowed to rise within the preset time window but should fall back to the baseline value within the preset recovery time. The system finds that the actual door opening time exceeds the preset standard opening time, and the Z3 particle concentration is above the preset threshold in two consecutive time windows. The power consumption of the fan filter unit-7 also rises simultaneously. After comparing the actual residual with the theoretical anomaly database, it is found that it is closest to the theoretical residual of personnel carrying materials into the area and prolonging the door opening, which leads to an increase in the pollution load in the core zone. Therefore, the system outputs the personnel violation type and automatically sends a retraining work order to the team leader and the terminal of the personnel involved after reaching the duration threshold. If it is found that the fan filter unit-7 maintains high power consumption and slow pressure recovery for a long time in the following period, the system can continue to identify the signs of equipment degradation in the subsequent time window and trigger the maintenance process on the equipment management end.

[0035] The purpose of this step is to transform cleanroom operation and maintenance from single-point threshold alarms to deviation diagnosis oriented to business scenarios. By first reconstructing the ideal baseline of the expected baseline operating state, and then aligning the actual deviations with theoretical anomalies, the background interference caused by production scheduling and normal logistics activities can be removed, and more accurate identification and handling of personnel violations, equipment degradation and space blockage can be achieved.

[0036] Constructing an ideal reference spatiotemporal tensor includes: parsing production scheduling data to obtain task time sequences; parsing personnel permissions and standard operating procedure data based on task time sequences to obtain standard behavior sequences; generating theoretical reference power consumption distribution, theoretical pressure difference distribution, and theoretical environmental load distribution based on task time sequences, standard behavior sequences, equipment ledgers and maintenance records, and spatial topology configuration data; and performing time alignment and spatial mapping on the theoretical reference power consumption distribution, theoretical pressure difference distribution, and theoretical environmental load distribution to obtain the ideal reference spatiotemporal tensor.

[0037] This embodiment provides a construction step for an ideal reference spatiotemporal tensor. Specifically, in the continuous scenario of the aforementioned aseptic preparation filling workshop, relying solely on historical averages will absorb past violations and equipment aging into the normal model, thus causing reference distortion. Therefore, this embodiment introduces an ideal reference reconstruction mechanism driven by production scheduling tasks, standard behaviors, equipment health status, and spatial topology.

[0038] Specifically, the system first parses the production scheduling data to obtain the task sequence; the task sequence can reflect the actual business rhythm of batch production. For example, T1 is the preparation for personnel to enter the site, T2 is the completion of sterile clothing changes and entry into the core filling area, T3 is the entry of the first batch of empty bottles, T4 is the start of filling, and T5 is the synchronous start of the capping process. The system does not regard all time slices as uniform and stable, but identifies the business actions corresponding to each time slice. Because in the cleanroom, access control opening, material flow, and equipment load changes have strong task dependence. The same particle fluctuation may be reasonable during material transfer, but may be abnormal during static pressure holding.

[0039] The system parses personnel permissions and standard operating procedure data based on task timing to obtain a standard behavior sequence. A standard behavior sequence refers to which type of personnel, via which route, and within what time period are allowed to complete what action at each task node. For example, liquid preparation personnel can only reach the pre-processing area, and filling operators can enter the core area after completing the prescribed changing process. Material transport personnel can only advance material carts through buffer paths, and access control should be opened in an interlocking sequence. This can be illustrated by a simplified application scenario: if at time T2, the standard route is Z1 changing area → Z2 buffer zone → Z3 core area, then the system binds this route and the corresponding access control event into a standard behavior chain. If a personnel's permissions only cover Z1 and Z2, then their standard behavior chain should not extend to Z3 at T2. The resulting sequence is not an abstract set of instructions, but rather a model of business constraints regarding who can do what, when, and where.

[0040] The system combines task timing, standard behavior sequences, equipment ledgers and maintenance records, and spatial topology configuration to generate theoretical baseline power consumption distribution, theoretical pressure difference distribution, and theoretical environmental load distribution. Among them, the theoretical baseline power consumption distribution reflects the minimum power consumption level required for ventilation and equipment operation in each area under the premise of meeting cleanliness level and process continuity; the theoretical pressure difference distribution reflects the target pressure difference gradient required to maintain the blockage of pollution from high-risk areas to low-risk areas.

[0041] The theoretical environmental load distribution reflects the particulate, heat, humidity, and disturbance loads that each area should bear under standard personnel numbers, standard door opening times, and standard logistics cycle times. Equipment maintenance records should also be considered here, because the health status and resistance of the same model of fan filter unit are different when it is close to the maintenance cycle compared to when it is just maintained. If this point is completely ignored, the benchmark will be out of touch with reality. Preferably, for equipment that has just completed maintenance, the rated health parameters in the log can be used as an ideal reference. For equipment that is close to the maintenance cycle but has not yet failed, a reasonable health reduction boundary can also be retained in the benchmark to avoid miscounting acceptable natural aging as part of the ideal state.

[0042] Subsequently, the system performs time alignment and spatial mapping on the three types of theoretical distributions mentioned above to form a unified ideal reference spatiotemporal tensor. Time alignment refers to unifying business cycles, equipment responses, and environmental changes from different sources under the same time slice. For example, when access control events are at the second level, energy consumption is at the minute level, and pressure difference is at the continuous sampling level, they are mapped to a common time window. Spatial mapping refers to projecting sensors, equipment, and business actions onto regional units. For example, P3, fan filter unit-7, and door G2 can all be mapped to relevant units in the Z3 core area, ultimately forming a comprehensive expression of what power consumption, pressure difference, and environmental load this area should have at time T3.

[0043] As a supplementary processing mechanism, if the production schedule is temporarily interrupted or the process sequence is adjusted, the system will prioritize using the latest issued task sequence to regenerate the baseline; if the personnel permission table is not updated in time, the personnel involved will only be modeled with conservative permissions and will not be inferred to have access to higher-level areas; if equipment maintenance records are missing, a temporary baseline will be established using the rated parameters of the equipment model and the conservative value of the most recent stable operation window, and marked as low confidence in the background; if a temporary partition is added to a certain area or the area name is changed, the area will not be included in the core scope of automatic diagnosis until the spatial mapping verification is passed.

[0044] For example, in the same batch of freeze-dried powder injection tasks, the system identifies from the manufacturing execution system that 9:00 AM to 9:30 AM is the core filling preparation stage, and 9:30 AM to 10:00 AM is the formal filling stage. According to the standard operating procedure, only two authorized operators can complete changing clothes and enter Z3 before 9:10 AM, and the opening of door G2 should be concentrated between 9:05 AM and 9:12 AM. During this stage, fan filter unit-7 and fan filter unit-8 should be gradually upgraded to maintain the state required for Class A laminar flow, but there should be no continuous overload. Based on this, the system forms different theoretical minimum energy consumption, pressure difference, and environmental load distributions at time slots such as 9:10 AM, 9:20 AM, and 9:40 AM, and maps them to areas Z1 to Z4. If the actual collected data deviates from these benchmarks, the system can determine whether the deviation exceeds the reasonable range of the current process stage.

[0045] The purpose of this step is to construct an ideal operational reference that is closely linked to specific batches, specific regions, and specific job actions, thereby realizing the transformation from historical average reference to business mechanism reference and providing a reliable foundation for subsequent abnormal residual extraction.

[0046] The abnormal event operators include time offset parameters, amplitude offset parameters, duration parameters, and spatial influence range parameters. The abnormal event operators are used to parametrically perturb the ideal reference spatiotemporal tensor, generating a set of abnormal simulation states. This includes: calling the expert rule base to determine personnel violation event operators, equipment degradation event operators, and logistics congestion event operators; using the personnel violation event operators to perform personnel behavior perturbations to obtain personnel abnormal simulation states; using the equipment degradation event operators to perform equipment performance perturbations to obtain equipment abnormal simulation states; and using the logistics congestion event operators to perform spatial obstruction perturbations to obtain flow field abnormal simulation states. Finally, the abnormal simulation states of personnel, equipment, and flow field are summarized to obtain the abnormal simulation state set.

[0047] This embodiment provides a parameterized perturbation mechanism for anomaly event operators. Specifically, in the case of only an ideal benchmark, although the system knows the expected benchmark operating state, it is still difficult to explain which type of anomaly it belongs to after deviation. Because different anomalies may cause particle rise or power fluctuation, if an interpretable anomaly simulation library is not introduced, the system will be unable to accurately distinguish different anomaly root causes. Therefore, this embodiment precipitates common cleanroom operation and maintenance problems as event operators and actively injects perturbations on the ideal benchmark to form comparable theoretical anomaly samples.

[0048] Specifically, the anomaly operator includes at least the time offset parameter, amplitude offset parameter, duration parameter, and spatial impact range parameter. The time offset parameter describes whether the anomaly occurs ahead of, synchronously with, or lagging behind the standard task node. For example, a logistics vehicle should enter the buffer zone at T3 but is stuck at the door at T2. The amplitude offset parameter describes the degree of deviation caused by the anomaly. For example, the increase in particle load caused by a door opening is greater than the preset normal load fluctuation threshold. In numerical processing, this is represented by applying a multiplicative gain or additive bias to the corresponding feature dimension of the ideal reference tensor. The duration parameter describes how long the anomaly effect lasts. For example, the slow recovery of pressure difference caused by door blockage may last for multiple time windows. In numerical processing, this is represented by stretching the above amplitude offset along the time dimension. The spatial impact range parameter describes which adjacent areas the anomaly extends from the source area. For example, personnel violations mainly affect the core area and its adjacent door areas after changing clothes, while logistics delays may spread along the corridor direction.

[0049] The system can call at least three types of event operators from the expert rule base. The first type is personnel violation event operators, which correspond to behaviors such as not following the changing procedure, entering without authorization, not closing the door in time, and multiple people tailgating into the room. The second type is equipment degradation event operators, which correspond to states such as dust accumulation on the filter screen of the fan filter unit, decreased fan efficiency, and drift of the differential pressure monitoring component. The third type is logistics congestion event operators, which correspond to situations such as material carts staying for too long, pass-through windows being occupied for too long, and turnover boxes being stacked to form local obstructions. The expert rule base can be jointly maintained by cleanroom operation and maintenance engineers, production managers, and validation engineers. Its content is not arbitrary experience, but is linked to process risk assessment, cleanroom validation conclusions, and equipment maintenance specifications.

[0050] When generating abnormal simulation states, the system executes different types of disturbances. For personnel behavior disturbances, actions such as unplanned door opening, personnel entry time delay, and entering the core area without completing the required stay can be inserted into the standard behavior chain in the ideal baseline. This changes the particle load and access control rhythm of the corresponding area. For equipment performance disturbances, without changing the production schedule, a certain equipment service area can exhibit a state of increased power consumption but slower cleanroom efficiency recovery or increased pressure difference fluctuation. For spatial obstruction disturbances, channel occupancy can be introduced on a specific path of topological connectivity. The consequence is not simply exceeding the standard at a certain point, but rather weakening the flow field locally, causing the pressure difference of adjacent doorways and buffer zones to be synchronously abnormal with the environmental load.

[0051] A simplified simulation scenario can be used for comparison and explanation. Assume that in the ideal baseline, there is a standard transfer from Z2 to Z3 at time T3, door G2 opens once and closes quickly, and the particle value of Z3 briefly rises and recovers. If a personnel violation event operator is introduced, an unplanned return door opening can be added after T3, resulting in two access control events in the actual monitoring data, interrupted particle fall, and extended compensation for the core area fan filter unit. If an equipment degradation event operator is introduced, the number of access control events is not increased, but the time for Z3 power consumption to exceed the preset baseline value after T3 is longer than the second preset time period, and the pressure difference recovery is slower than the ideal baseline. If a logistics stagnation event operator is introduced, the access control may be normal, but there will be continuous pressure difference disturbances in the area between Z2 and Z3, spreading to neighboring areas. The resulting theoretical anomalies are distinguishable from each other, and subsequent comparisons will no longer rely solely on a single indicator.

[0052] In further boundary condition configuration rules, if there is no event template corresponding to the current new equipment or process in the expert rule base, the system can first adopt a general disturbance framework, retain only the four basic parameters of time, amplitude, duration and space, generate a low-precision simulation state, and mark the anomaly as a type to be expanded; if multiple anomalies can occur simultaneously, such as personnel violations superimposed on equipment degradation, a combined anomaly simulation state can be generated to identify composite root causes; if the application of an event operator causes the simulation result to violate the basic engineering constraints of the cleanroom, such as the pressure difference direction being completely opposite to the closed area design, then the simulation state is regarded as an invalid template and will not be included in the theoretical anomaly database.

[0053] For example, in the aforementioned filling workshop, the system constructs simulations by calling three types of event operators to investigate anomalies after the start of filling in the morning. For personnel violations, it simulates an operator opening the G2 door again shortly after material transfer to return to the buffer zone to retrieve consumables. For equipment degradation, it simulates a decrease in the core area's airflow maintenance capacity after the resistance of the filter material in the fan filter unit -7 increases. For logistics congestion, it simulates a material truck stopping at the intersection of Z2 and Z3 for more than the standard waiting time. Although all three simulation states may cause particle and power consumption anomalies, they have different multidimensional data fluctuation characteristics in terms of the order of occurrence, the area of ​​influence, and the mode of duration, thus providing a more interpretable theoretical comparison reference for subsequent diagnosis.

[0054] The purpose of this step is to transform common management and equipment problems in cleanrooms into a reproducible and comparable database of abnormal patterns, thereby enabling a refined identification of the root causes of actual deviations, rather than remaining at the level of coarse-grained alarms where there are abnormalities but the cause is unknown.

[0055] The actual residual matrix is ​​a deviation matrix with time slices as rows and preset monitoring feature dimensions as columns. The preset monitoring feature dimensions include at least one of the following: area power consumption deviation, access control trigger count deviation, door opening duration deviation, particle concentration deviation, pressure difference deviation, and environmental recovery time deviation.

[0056] Each theoretical residual matrix in the theoretical residual matrix set has the same dimension as the actual residual matrix; the residuals in the first and second differences are absolute differences, relative deviations, or normalized differences; before generating diagnostic results, the following steps are also included: normalizing the actual residual matrix and the theoretical residual matrix set and splicing them according to time windows to construct a unified feature space.

[0057] This embodiment provides a mechanism for constructing residual matrices and organizing unified feature spaces. Specifically, in the aforementioned scenario, even with ideal benchmarks and abnormal simulation states, if the feature dimensions used in the real data and theoretical data are inconsistent, or if the data dimensions of features such as power consumption, access control, particle concentration, and pressure difference are not uniform, subsequent comparisons will be biased towards features with large numerical ranges, making it difficult to reflect the true meaning of operation and maintenance. Therefore, this embodiment uses unified dimensions, unified time windows, and unified feature scales to provide a basis for direct comparison between real and theoretical deviations.

[0058] Specifically, the real residual matrix is ​​arranged with time slices as rows and monitoring feature dimensions as columns. Optional feature dimensions include regional power consumption deviation, access control trigger count deviation, door opening duration deviation, particle concentration deviation, pressure difference deviation, and environmental recovery time deviation. For example, in a simplified verification model, with T1, T2, and T3 as three rows and F1 power consumption, F2 access control, F3 particles, and F4 pressure difference as four columns, the real residual matrix can be represented as a 3×4 deviation representation structure. Correspondingly, each theoretical anomaly simulation also forms a 3×4 theoretical residual matrix. In this way, regardless of the similarity method used for subsequent comparison, both are expressed in the same coordinate system.

[0059] The residuals in the first and second differences can be expressed as absolute differences, relative deviations, or normalized differences. Absolute differences are suitable for expressing indicators with clear engineering upper limits, such as how many seconds the door opening time deviates from the standard value. Relative deviations are suitable for expressing health deviations when different equipment has inconsistent rated values. For example, the power consumption deviation of two fan filter units with different power ratings should not be directly compared in wattage, but rather the proportion of deviation relative to the rated operating level should be considered. Normalized differences are suitable for multi-source fusion scenarios, enabling particle, pressure difference, and power consumption to be compared within a unified range. Preferably, different difference methods can be used for different characteristics, and then uniformly mapped to a standard range to retain their respective engineering significance.

[0060] Before generating diagnostic results, the system normalizes and splices the actual and theoretical residual matrices by time windows to construct a unified feature space. Normalization is not simply about compressing values, but about reducing dimensional differences, allowing for a slight increase in power consumption that lasts longer and a sharp rise in particles that recovers quickly, all within the same comparative framework. Time window splicing is used to preserve the continuity of the abnormal process, because cleanroom anomalies are often not isolated points, but a series of interconnected dynamic events. For example, a violation of door opening rules can manifest as a first access control deviation, then a particle deviation, and then a power consumption compensation deviation. If only a single time slice is considered, the causal chain will be broken. By splicing by time windows, the system can combine several adjacent time slices to observe the evolution of the deviation.

[0061] A simplified example can be used to illustrate this; assuming that in reality, the access control deviation in three consecutive time slices is none, strong, and weak in sequence, the particle deviation is weak, strong, and strong in sequence, and the differential pressure deviation is none, weak, and weak in sequence, then after splicing, a continuous characteristic segment is formed that reflects the rise of particles after the door is opened, accompanied by differential pressure disturbance; if a theoretical anomaly also presents a similar sequence, then the two are more likely to belong to the same root cause; conversely, if a theoretical anomaly presents that the power consumption recovery time is greater than the preset recovery time threshold, and then the differential pressure continues to weaken while the access control remains unchanged, then it is closer to equipment degradation than personnel violation;

[0062] As a supplementary processing mechanism, if a certain feature is missing in some time windows, such as when a particle sensor is temporarily interrupted during calibration, it can be temporarily filled by using redundant sensors in the same area or steady-state values ​​from adjacent time windows. However, the weight of the filled feature needs to be reduced. If all key features are incomplete within a certain time window, they will not participate in automatic diagnosis, and only observation records will be retained. If a certain type of feature remains in a zero-fluctuation state for a long time, such as when a backup door has never been used in the current batch, the feature dimension can be removed from the current diagnostic window to avoid introducing invalid comparisons.

[0063] For example, during the anomaly analysis of the filling core area, the system selects four consecutive 5-minute time slices and uses the power consumption deviation of the fan filter unit -7, the door opening deviation of G2, the particle deviation of P3, and the pressure difference deviation of Z2-Z3 as four feature dimensions. Actual data shows that the door opening deviation is the most prominent in the second time slice, the particle deviation reaches its peak in the third time slice, and the power consumption is still higher than the baseline in the fourth time slice. The system performs unified normalization and splicing comparison of this actual residual segment with multiple theoretical residual segments to identify an anomaly chain that is closer to the personnel returning to open the door, rather than a long-term deviation chain of slow degradation of the fan filter unit.

[0064] The purpose of this step is to ensure that real-world anomalies and theoretical anomalies are compared under the same structure, thereby enhancing the comparability of multi-source heterogeneous data and avoiding misjudgments caused by differences in units, sampling frequency and missing data.

[0065] Based on the temporal and spatial similarity between the actual residual matrix and the theoretical residual matrix set, a diagnostic result is generated, including: calculating the temporal similarity between the actual residual matrix and each theoretical residual matrix using dynamic time warping; calculating the spatial similarity between the actual residual matrix and each theoretical residual matrix using cosine similarity; weighting and fusing the temporal and spatial similarities according to preset weights to obtain a comprehensive matching score; comparing the comprehensive matching score with preset high and low thresholds, where the preset high threshold is greater than the preset low threshold; outputting the corresponding abnormality type when the comprehensive matching score is greater than or equal to the preset high threshold; outputting a normal result when the comprehensive matching score is less than or equal to the preset low threshold; and outputting a review result when the comprehensive matching score is greater than the preset low threshold and less than the preset high threshold.

[0066] This embodiment provides a diagnostic result generation mechanism based on temporal and spatial similarity. Specifically, after the aforementioned residual matrix has been unified, if only the numerical value of the anomaly is considered without considering its order and diffusion location, different root causes may still be confused. For example, both equipment degradation and personnel violations may cause particle increases, but the former often exhibits a state evolution where the parameter change rate is less than a preset change rate threshold and the duration is greater than a preset duration threshold, while the latter is usually accompanied by access control events and concentrated in local areas. Therefore, this embodiment incorporates both temporal evolution characteristics and spatial distribution characteristics into the diagnosis.

[0067] Specifically, the system uses dynamic time warping to calculate the temporal similarity between the actual residual matrix and each theoretical residual matrix. Its engineering implications are that even if there is a slight difference in the exact timing of actual anomalies and theoretical templates, as long as their change rhythms are similar, they can still be identified as similar anomalies. This is particularly important for cleanrooms because different shifts performing the same procedures may experience natural fluctuations of a few minutes during door opening, transfer, and environmental restoration. If strict time coincidence is required, the system will be overly sensitive to normal operational fluctuations. Dynamic time warping, in essence, compares how anomalies unfold while tolerating reasonable time offsets.

[0068] Specifically, the system calculates the dynamic time-warped cumulative shortest path distance between the time-series characteristic sequences of the actual residual matrix and the time-series characteristic sequences of the theoretical residual matrix, using a negative exponential mapping, for example... Where St is the temporal similarity score, DTW is the dynamic time warping cumulative shortest path distance, α is a positive hyperparameter controlling the steepness of the mapping; the value of the positive hyperparameter α can be calibrated based on the historical steady-state test data of the target cleanroom to match the specific spatial scale, and e is the base of the natural logarithm; or the inverse proportional mapping can convert the DTW distance into a temporal similarity score falling in the [0,1] interval, therefore The smaller the distance, the closer the corresponding temporal similarity score is to 1;

[0069] Meanwhile, the system uses cosine similarity to calculate the spatial similarity between the actual residuals and each theoretical residual. Specifically, the system flattens the residual features of each region within the decision time window into a one-dimensional spatial representation vector according to a fixed spatial topological order, calculates the dot product of the actual residual spatial vector and the theoretical residual spatial vector, and divides it by the product of their magnitudes to obtain the cosine similarity score. Since dimensional constraints have been applied before the difference extraction, the consistency of spatial direction can reliably fall within a unified interval.

[0070] The engineering implications are that even if the total amount of anomalies is similar, if they mainly occur in different areas, the root causes may still be different. For example, pollution loads caused by personnel violations usually spread around personnel entry paths and doorway areas, while equipment degradation is more likely to manifest as continuous environmental deterioration around specific equipment service areas. By comparing the consistency of the residual directions in each area, the system can determine whether the actual anomaly spreads along the same spatial contour as the theoretical anomaly.

[0071] The system then weights and fuses the temporal and spatial similarity scores according to preset weights to obtain a comprehensive matching score that falls within a unified range. For access control-driven anomalies, temporal evolution is often more identifiable, so the temporal weight can be appropriately increased. For device degradation in fixed areas, spatial contours are often more stable, so the spatial weight can be appropriately increased. The weights can be initialized based on expert experience or updated by subsequent feedback. In the specific calculation, the normalized temporal similarity score is set as... Spatial similarity score is The corresponding temporal weights and spatial weights are respectively and and satisfy Then the overall matching score It can be represented as:

[0072]

[0073] A simplified quantization example can be used to illustrate the fusion calculation process: if we set... , The temporal similarity between the actual residual and the personnel violation template A was measured to be: Spatial similarity is The overall matching score is calculated as follows:

[0074]

[0075] Assuming the high threshold is set to 0.85 and the low threshold is set to 0.60, since 0.87 is greater than or equal to the high threshold, the fused comprehensive matching score enters the high range, and the corresponding abnormal type is directly output; if the calculated comprehensive matching score is less than or equal to 0.60, a normal result is output; if the comprehensive matching score is 0.75, which is in the middle range of greater than the low threshold and less than the high threshold, a review result is output.

[0076] Regarding the judgment interval, the high threshold is greater than the low threshold. When the comprehensive matching score is not lower than the high threshold, it indicates that the actual deviation and a certain theoretical anomaly have largely overlapped in time and space, and the corresponding anomaly type can be directly output. When the comprehensive matching score is not higher than the low threshold, it indicates that the actual deviation is still within the normal process fluctuation range, or that the existing theoretical templates cannot explain the deviation, and a normal result can be output. When it is between the two, it indicates that the existing evidence is insufficient to support automatic qualitative analysis, and it needs to be transferred to the review process. This dual threshold design can prevent the system from directly outputting a mandatory diagnostic conclusion in the boundary state of insufficient evidence.

[0077] In the handling logic when multiple templates conflict in abnormal features, if the comprehensive matching scores of two different abnormal types both exceed the high threshold, the consistency of their key features can be further compared, such as whether there is an access control event as a trigger source, or whether there is a long-term high power consumption as a clue to health degradation; if they still cannot be distinguished, the verification result is output instead of forced merging; if the actual residual is in the middle range for all theoretical templates, it may mean that a new abnormal type or multiple abnormalities have appeared, and the system can establish a temporary case and enter the manual annotation process; if the spatial topology has been temporarily adjusted in the current shift, the spatial similarity weight can be appropriately reduced to avoid the old topology interfering with the judgment;

[0078] For example, in an anomaly in the aforementioned filling workshop, the actual residuals show: first, an abnormal opening time of door G2 occurred, followed by an increase in particle concentration in Z3, and then a short-term power consumption compensation for the fan filter unit -7. Simultaneously, the deviations were mainly concentrated at the Z2-Z3 junction and the entrance to the Z3 core area. After comparing these residuals with the theoretical residuals for personnel violations, equipment degradation, and logistics delays, the system found that the personnel violation template was closest in time sequence and its spatial distribution was concentrated at the entrance and core area. The overall matching score exceeded the high threshold, so the personnel violation result was directly output. In another case, only the power consumption recovery time of the fan filter unit -7 was greater than the preset recovery time threshold, and the Z3 pressure difference gradually worsened, while the access control and logistics rhythm were basically normal. In this case, the equipment degradation template achieved a higher match. If the data in an anomaly falls between the two templates, the system outputs a verification result, which is then confirmed by the maintenance engineer through video and on-site inspection.

[0079] The purpose of this step is to incorporate both when and where the abnormality occurs into the judgment logic, and to eliminate the dimensional differences between distance calculation and similarity assessment, thereby enabling interpretable diagnosis of complex deviations in cleanrooms and reducing false alarms and mixed reports caused by the similarity of a single indicator.

[0080] The preset dispatch conditions include at least one or more of the following: anomaly type, comprehensive matching score threshold, anomaly duration threshold, and impact level threshold; generate maintenance work orders and send handling instructions, including: when the anomaly type is personnel violation, generate a retraining work order and send it to the corresponding personnel's terminal; when the anomaly type is equipment degradation, generate a maintenance work order and send it to the equipment management terminal; when the anomaly type is spatial obstruction, generate a dispatch work order and send it to the logistics management terminal; and associate and store the retraining work order, maintenance work order, or dispatch work order with the diagnostic results.

[0081] This embodiment provides a linkage mechanism for work order dispatch and handling. Specifically, after the aforementioned diagnostic results have been formed, if the system only stays at the level of discovering anomalies and cannot transform the anomaly type into handling actions for specific responsible parties, the diagnostic effectiveness will be difficult to close the loop, and resource misallocation will easily occur, where all anomalies are handled centrally and uniformly by facility personnel. Therefore, this embodiment uses anomaly type, matching degree, duration, and impact level together for work order dispatch.

[0082] Specifically, the preset dispatch conditions may include at least one or more of the following: anomaly type, comprehensive matching score threshold, anomaly duration threshold, and impact level threshold. The anomaly type is used to determine who is subject to handling; the comprehensive matching score is used to determine whether the evidence for automatic dispatch is sufficient; the duration is used to distinguish between transient disturbances and persistent problems; and the impact level is used to distinguish whether it affects core process areas or critical batches. For cleanrooms, this combination of multiple conditions is essential because not all anomalies should immediately interrupt production or issue mandatory work orders. For example, if a short-term door opening is prolonged but does not lead to a continuous deterioration of environmental indicators in the core area, it is more appropriate to record and issue a warning first. However, if the anomaly continues to span multiple time windows and affects Class A or Class B areas, automatic dispatch should be prioritized.

[0083] In terms of handling actions, when the anomaly type is personnel violation, the system generates a retraining work order and sends it to the corresponding personnel terminal or team leader terminal. The retraining work order can include the time of the anomaly, the corresponding area, the deviated SOP node, and suggested review items, such as changing procedures, access control operation specifications, and return retrieval requirements. When the anomaly type is equipment degradation, the system generates a maintenance work order and sends it to the equipment management terminal. The work order content can be associated with the specific equipment number, service area, anomaly manifestation, most recent maintenance time, and suggested inspection direction, such as filter material resistance, fan operating efficiency, or differential pressure sampling component status. When the anomaly type is spatial obstruction, the system generates a scheduling work order and sends it to the logistics management terminal, prompting the replanning of transfer routes, clearing of obstructing materials, or adjustment of queuing order. All of the above work orders are stored in association with the diagnostic results to maintain the entire evidence chain from the discovery of deviations to the execution of actions.

[0084] A simplified multi-condition judgment structure tree example can be used to illustrate the work order triggering rules. If the comprehensive matching score of a personnel violation anomaly is greater than or equal to the preset high threshold, lasts for more than two time windows, and affects the core area entrance, a retraining work order is immediately dispatched. If the comprehensive matching score of an equipment degradation anomaly is between the preset low threshold and high threshold, and the duration is greater than the second preset duration threshold, and the equipment is close to its maintenance cycle, the dispatching conditions can also be met. If a spatial blockage anomaly only occurs briefly in the buffer zone and does not affect the core area, an observation record is first created. If a similar blockage occurs again, it is upgraded to a scheduling work order. It can be seen that the dispatching is not a fixed logic that directly triggers instruction distribution based on a single variable limit, but rather a hierarchical response and handling based on multi-dimensional business characteristics.

[0085] Regarding the specific handling methods for situations where the responsible party is unreachable or complex anomalies, if the anomaly type is clear but the responsible party is temporarily unreachable (e.g., the corresponding personnel's terminal is offline), the work order can be copied to both the team leader and the quality manager. If the equipment management terminal does not respond within the specified time, the work order is escalated to the facility supervisor. If the same event involves both personnel violations and equipment degradation, the system can generate a master work order and related sub-work orders, targeting different responsible departments respectively, to avoid shirking responsibility or hindering business processes. If the diagnostic results are in the review period, a mandatory work order is not generated directly, but a review task is first sent to the on-duty maintenance terminal. If the current batch is in a critical period before release, the system can also associate relevant anomaly work orders with batch records for review during quality audits.

[0086] For example, shortly after filling begins, the system determines that an anomaly closely matches the theoretical template of personnel returning to open the door, and that the anomaly continues to affect the entrance of the Z3 core area for two consecutive time windows. Therefore, it automatically generates a retraining work order, pushes it to the affected operator and team leader, and records the access control logs and environmental deviation screenshots for the corresponding time slot. During another period, the system identifies that the fan filter unit-7 has persistent high power consumption and slow differential pressure recovery. Based on this, it determines that the equipment has reached the degradation dispatch threshold and generates a maintenance work order, sending it to the equipment management terminal, requiring priority inspection of the filter and fan status after this batch is completed. If material carts are found to be repeatedly stuck at the buffer zone entrance during peak material transfer periods, the system generates a scheduling work order and sends it to the logistics management terminal, requiring adjustments to the transfer rhythm to reduce congestion at the entrance.

[0087] The purpose of this step is to directly map diagnostic conclusions to maintenance actions that match the responsible parties, thereby achieving business collaboration in anomaly identification, responsibility distribution, and implementation confirmation, and solving the maintenance gap problem where single-point facility alarms are disconnected from on-site repair actions.

[0088] The perturbation parameters of the abnormal event operator and the matching score threshold or duration threshold in the preset distribution conditions are updated based on the judgment accuracy index. This includes: when the judgment accuracy index is lower than the preset accuracy threshold, adjusting the time offset parameter, amplitude offset parameter, duration parameter or spatial influence range parameter of the abnormal event operator, and adjusting the matching score threshold or duration threshold in the preset distribution conditions; when the judgment accuracy index is higher than or equal to the preset accuracy threshold, retaining the current parameters and thresholds.

[0089] This embodiment provides an evaluation and update mechanism based on execution feedback. Specifically, in cleanroom operation and maintenance, even if the initial rules and anomaly templates are derived from expert experience, it is difficult to fully adapt them to all workshops, work groups, and equipment statuses at once. If the system maintains fixed disturbance parameters and fixed dispatch thresholds for a long time, false alarms or missed alarms are likely to increase after process adjustments, changes in personnel proficiency, or equipment aging phases. Therefore, this embodiment uses work order execution results and review results to reverse-correct the anomaly templates and dispatch thresholds.

[0090] Specifically, the system generates a judgment accuracy index based on the execution result data and review result data of the maintenance work order. The execution result data may include whether the equipment power consumption has recovered after maintenance, whether the differential pressure has recovered to the design gradient, whether the doorway obstruction has been eliminated after logistics adjustments, and whether the same deviation reappears in subsequent shifts after personnel retraining. The review result data may include the on-site verification conclusions of the quality department, the video review conclusions, and the actual anomaly types confirmed by manual verification. The judgment accuracy index can be understood as a comprehensive measure of whether the system's diagnosis has been verified by subsequent facts.

[0091] In the specific calculation, the system records work order records with consistent review results and positive execution results (i.e., the corresponding handling is effective and the actual abnormal characteristics have obviously declined) as positive judgment samples according to a preset verification cycle. Records with inconsistent root causes found during review or records that remain unchanged after handling are recorded as negative judgment samples. The system weights the positive judgment samples generated by the specified type within the statistical window period and divides it by the weighted sum of the total samples triggered by the corresponding type to quantify and output the percentage of judgment accuracy of each event operator.

[0092] Furthermore, if the first The consistency results of each sample are denoted as The value is set to 1 if the verification is consistent, and 0 otherwise; and the risk weight of the spatial region where the sample is located is denoted as Then, within the statistical window period, the accuracy index for judging this type of anomaly... It can be represented as:

[0093]

[0094] Among them, in the numerator and denominator The total number of samples that triggered this type of exception; For the first Risk weights for the spatial regions where each triggering sample is located. The summation symbol;

[0095] If different regions have different impact levels, risk weights can be applied to make the erroneous judgments in the core region have a greater penalty effect in the denominator compared to ordinary regions. If the system uses the above method to count that a certain event has repeatedly misjudged normal transportation as spatial blockage, it means that the spatial impact range parameter setting value of the corresponding event operator is greater than the preset upper limit, or the dispatch threshold parameter setting value is less than the preset lower limit. If the system fails to identify early equipment degradation in time, and the equipment is subsequently diagnosed with filter blockage, it means that the duration or amplitude parameter of the equipment degradation event operator is lower than the tolerance range required for actual operation.

[0096] When the accuracy index of the judgment is lower than the preset accuracy threshold, the system adjusts the time offset parameter, amplitude offset parameter, duration parameter or spatial influence range parameter of the abnormal event operator, and adjusts the matching score threshold or duration threshold accordingly; the update here is not an arbitrary change, but a targeted correction along the source of error;

[0097] For example, if manual verification confirms that a certain type of personnel violation usually occurs shortly after the material transfer is completed, but the current template sets it before the transfer, the time offset parameter should be adjusted; if the actual power consumption change in the equipment degradation case does not exceed the preset amplitude threshold, but the differential pressure recovery time is consistently greater than the preset recovery time threshold, the power consumption amplitude feature should be weakened and the duration feature strengthened; if too many minor deviations are automatically assigned, the matching score threshold can be increased or the duration threshold extended; conversely, when the judgment accuracy index reaches or exceeds the preset accuracy threshold, the current parameters and thresholds should be retained to maintain model stability.

[0098] A simplified verification model example can be used for illustration; if, in the most recent 10 spatial obstruction judgments, 4 are actually normal peak transfers after manual review, then the accuracy of this type of diagnosis is insufficient. The system can reduce the spatial impact range of the logistics congestion event operator, requiring it to affect the area near the entrance and last for a longer period before triggering a dispatch order; if 8 out of the most recent 10 equipment degradation judgments are confirmed after maintenance, it indicates that the template is relatively reliable and can be temporarily left unupdated; through this result-based structured statistical fine-tuning, the system will gradually converge from general rules to dedicated rules that better fit the operating habits of this workshop;

[0099] As a supplementary processing mechanism, if the feedback data itself is incomplete, such as the work order being closed but the maintenance results not being entered, or the manual review not reaching a conclusion, then the case will not participate in parameter updates and will only be retained as a sample to be supplemented; if a large-scale process change occurs in the short term, such as adding a new filling specification or adjusting the purpose of a clean area, the automatic update of historical parameters can be temporarily frozen and reactivated after the new process is running stably; if the abnormal change rate of the number of abnormal alarms exceeds the preset mutation threshold after an update, the system can roll back to the previous stable parameter version and prompt the operation and maintenance administrator to review the update strategy;

[0100] For example, after a month of continuous operation, the system found that a significant portion of the space obstruction work orders, after review by the logistics supervisor, were determined to be normal batch transfers, involving only short-term queues at the entrance and no continuous obstruction. Furthermore, the statistics module calculated that the accuracy rate of this abnormal dispatch rule was lower than the preset available accuracy rate threshold. Therefore, the system narrowed the spatial impact range parameter for this type of event, no longer identifying it as an obstruction based solely on single-point pressure difference fluctuations in the buffer zone, and simultaneously raised the duration threshold. On the other hand, after maintenance, the equipment degradation work orders were repeatedly confirmed to have dust accumulation on the fan filter unit screen, and the anomaly often manifested as initially maintaining high power consumption followed by a slower recovery of pressure difference. Moreover, the quantitative judgment accuracy rate remained consistently above 90%. Therefore, the system retained the core parameters of this type of template. After the update, the number of mis-dispatched orders in subsequent shifts significantly decreased, while cases that truly required equipment maintenance could still be identified in a timely manner.

[0101] The purpose of this step is to reverse-structure and correct the diagnostic template and dispatch rules based on the handling results, so as to achieve adaptive optimization of the system in long-term operation, taking into account both diagnostic sensitivity and business acceptability.

[0102] Example 2:

[0103] Please see Figure 2 An intelligent operation and maintenance system for cleanrooms based on artificial intelligence includes a processor and a memory. The memory is coupled to the processor and is used to store computer program code, which includes computer instructions. The processor calls the computer instructions to cause the system to perform the method as described in the embodiment.

[0104] This embodiment provides an intelligent operation and maintenance system for cleanrooms based on artificial intelligence. Specifically, the system can be deployed in the industrial server, private cloud platform, or edge computing gateway of the aforementioned aseptic preparation filling workshop. It adopts a collaborative architecture of processor and memory to execute the data acquisition, benchmark reconstruction, anomaly simulation, residual analysis, diagnostic dispatch, and feedback update processes in the aforementioned embodiments.

[0105] Specifically, the processor can be a general-purpose industrial central processing unit, a computing module with parallel processing capabilities, or a combination of both; the memory can include volatile and non-volatile memory, used to store program instructions, spatial topology configuration, expert rule base, historical work order records, equipment ledger, production scheduling snapshots, and model parameters; the system is connected to access controllers, environmental sensors, fan filter unit controllers, manufacturing execution system servers, and equipment management systems via industrial Ethernet or fieldbus; after the processor calls the program stored in the memory, it can first pull business and IoT data from different interfaces, and then organize various types of data according to time windows and regional units to form an ideal reference spatiotemporal tensor, an abnormal simulation state set, and a residual matrix, and output diagnostic results and work order instructions;

[0106] In terms of implementation, the system can be centrally deployed in a plant-level data center, managing multiple cleanrooms from a unified platform; or it can be distributed and deployed at the edge of a single cleanroom to perform near real-time diagnostics on access control, differential pressure, and power consumption of fan filter units, and then transmit the results back to the upper-level management platform. For alarms and dispatching actions with high latency requirements, it is preferable to complete them at data processing nodes close to the cleanroom. Parameter updates, rule maintenance, and statistical analysis can be performed on the central platform. Through this hardware and software combined architecture, the system has both real-time performance and cross-batch and cross-cleanroom management capabilities.

[0107] As a supplementary processing mechanism, if communication between the central platform and the workshop edge nodes is interrupted, the edge nodes can still use the latest topology, production schedule and rules cached locally to perform basic diagnostics, but large-scale parameter updates will be suspended; if the expert rule base version in the memory is abnormal or corrupted, the system can roll back to the most recently verified stable version; if the processor load is too high and cannot complete all simulation comparisons, the core area related diagnostic tasks will be retained first, and non-critical areas will be transferred to the waiting queue.

[0108] For example, in this filling workshop, the industrial server stores the topology of areas Z1 to Z4, the health records of each fan filter unit, and the current batch production schedule information; the processor periodically reads the access control log, P3 particle concentration, and fan filter unit -7 power consumption, and executes the aforementioned ideal benchmark reconstruction and theoretical anomaly simulation process; when a personnel violation anomaly is detected and the dispatch conditions are met, the system immediately pushes a retraining work order to the personnel terminal through the internal network; after the batch is completed, the system summarizes the work order execution results and updates the relevant event operator parameters;

[0109] The purpose of this step is to support the aforementioned operation and maintenance methods in the form of a deployable hardware and software system, so that it can continuously and stably perform intelligent diagnosis and closed-loop management functions in a real cleanroom environment.

[0110] The system includes: a data acquisition module for collecting business data sources, IoT data sources, and spatial topology configuration data; a benchmark reconstruction module for constructing an ideal benchmark spatiotemporal tensor; an anomaly perturbation module for generating a set of anomaly simulation states; a difference module for generating sets of real and theoretical residual matrices; a decision module for generating diagnostic results; a work order dispatch module for generating maintenance work orders and sending disposal instructions; a feedback acquisition module for collecting execution result data and review result data; and an evaluation and update module for generating decision accuracy indicators and updating the perturbation parameters of the anomaly event operator as well as the matching score threshold or duration threshold in the preset dispatch conditions.

[0111] This embodiment provides a modular intelligent operation and maintenance system structure for cleanrooms. Specifically, to facilitate engineering deployment, fault isolation, and functional expansion, the entire system can be divided into a data acquisition module, a baseline reconstruction module, an anomaly disturbance module, a differential module, a decision module, a work order dispatch module, a feedback acquisition module, and an evaluation and update module. Each module can be deployed on the same server or distributed across edge nodes and the central platform to operate collaboratively.

[0112] Specifically, the data acquisition module gathers business data sources, IoT data sources, and spatial topology configuration data. It can integrate multiple protocol adapters to interface with MES, access control controllers, sensor gateways, and equipment management systems, and performs unified correction on timestamps. The benchmark reconstruction module receives the data output by the acquisition module and forms an ideal benchmark spatiotemporal tensor corresponding to the current batch of tasks. The anomaly disturbance module generates a set of abnormal simulation states such as personnel violations, equipment degradation, and logistics delays on the ideal benchmark based on an expert rule base and parameter templates. The difference module is responsible for aligning and comparing the real data with the ideal benchmark and the theoretical anomalies with the ideal benchmark, generating sets of real residual matrices and theoretical residual matrices. The decision module performs temporal and spatial dual-dimensional matching and outputs the anomaly type, comprehensive matching score, and conclusions such as normal, review, or anomaly.

[0113] The work order dispatch module receives the results from the judgment module and generates work orders and issues processing instructions based on the dispatch conditions; the feedback collection module continuously tracks the work order execution process, receiving equipment maintenance records, personnel retraining feedback, logistics adjustment confirmations, and manual review conclusions; the evaluation and update module forms judgment accuracy indicators based on feedback and adjusts the abnormal event operator parameters and dispatch thresholds; the modules can transmit status through a message bus or shared database to ensure data link traceability.

[0114] A simplified module workflow example can be adopted: the data acquisition module acquires relevant data of the Z3 core area from T1 to T3; the benchmark reconstruction module forms the ideal benchmark for this time window; the anomaly disturbance module outputs three types of theoretical anomaly templates; the difference module unifies the actual deviation with the three types of templates; the decision module identifies the closest person as violating the rules; the work order dispatch module generates a retraining work order based on this; the feedback acquisition module records that the team leader has completed the retraining confirmation; the evaluation and update module finds that the accuracy rate of this type of work order review is high, so it retains the current parameters; through this hierarchical organization, the system is easier to maintain and expand. For example, when adding a new anomaly template with a continuously open transmission window, only the local rules of the anomaly disturbance module and the decision module need to be expanded.

[0115] As a supplementary processing mechanism, if a module malfunctions, a module-level degradation strategy can be adopted. For example, if the data acquisition module partially fails, it can be maintained for a short period of time by cached data; if the abnormal disturbance module is unavailable, the system can suspend root cause classification and only retain observation and early warning based on baseline deviation; if the work order dispatch module is unreachable, the judgment result is first written to the pending queue and resent after communication is restored; if the evaluation and update module is suspended, the existing stable parameters can still support daily diagnosis and avoid the impact of abnormal learning links on basic operation and maintenance.

[0116] For example, during a complete production day in a sterile preparation workshop, the data acquisition module continuously acquires production scheduling, access control, and environmental data; the baseline reconstruction module establishes ideal baselines for the morning filling and afternoon cleaning phases respectively; the anomaly disturbance module generates personnel violation templates applicable to the filling phase and logistics blockage templates applicable to the cleaning phase based on the rule base; the differential and decision module identifies a violation by personnel at the core area entrance during the noon period, and the work order dispatch module issues a retraining work order to the relevant personnel terminal; that evening, the feedback acquisition module receives the retraining completion record and quality review comments, and the evaluation update module updates the anomaly duration threshold for the team in the entrance area accordingly; the entire module link unfolds continuously around the same workshop, the same batch, and the same operation and maintenance closed loop;

[0117] The purpose of this step is to break down the complex cleanroom intelligent operation and maintenance process into implementable, maintainable, and scalable functional units through a modular architecture, thereby achieving the practical deployment and long-term stable operation of the system engineering.

[0118] 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 method for intelligent operation and maintenance of cleanrooms based on artificial intelligence, characterized in that, include: Collect business data sources, IoT data sources and spatial topology configuration data. Spatial topology configuration data includes at least the area division, area connectivity, sensor point-to-area mapping relationship and device-to-area mapping relationship. Based on production scheduling data, personnel access and standard operating procedure data, equipment ledger and maintenance record data, and spatial topology configuration data, an ideal reference spatiotemporal tensor organized according to preset time granularity and spatial partitioning is constructed. The ideal reference spacetime tensor is parametrically perturbed using anomaly event operators to generate a set of anomalous simulation states; The device power consumption data, access control trigger data, and environmental monitoring data are first differencing the corresponding spatial slices in the ideal reference spatiotemporal tensor to obtain the actual residual matrix. The abnormal simulation states are then second differencing the ideal reference spatiotemporal tensor to obtain the theoretical residual matrix set. Based on the temporal and spatial similarity between the real residual matrix and the theoretical residual matrix set, diagnostic results are generated, which include the anomaly type and the comprehensive matching score. When the diagnosis result meets the preset dispatch conditions, an operation and maintenance work order is generated and a disposal instruction is sent; otherwise, an observation record is generated. The preset dispatch conditions include at least a comprehensive matching score threshold or an abnormal duration threshold. Collect execution result data and review result data of operation and maintenance work orders, generate judgment accuracy index, and update the disturbance parameters of abnormal event operators and the comprehensive matching score threshold or abnormal duration threshold in the preset dispatch conditions based on the judgment accuracy index. Based on the temporal and spatial similarity between the actual residual matrix and the theoretical residual matrix set, diagnostic results are generated, including: Dynamic time warping is used to calculate the temporal similarity between the actual residual matrix and each theoretical residual matrix; Cosine similarity is used to calculate the spatial similarity between the actual residual matrix and each theoretical residual matrix; Temporal similarity and spatial similarity are weighted and fused according to preset weights to obtain a comprehensive matching score. The comprehensive matching score is compared with preset high threshold and preset low threshold respectively, where the preset high threshold is greater than the preset low threshold. When the comprehensive matching score is greater than or equal to the preset high threshold, the corresponding anomaly type is output; when the comprehensive matching score is less than or equal to the preset low threshold, a normal result is output; when the comprehensive matching score is greater than the preset low threshold and less than the preset high threshold, a review result is output. The preset distribution conditions include at least one or more of the following: anomaly type, comprehensive matching score threshold, anomaly duration threshold, and impact level threshold; Generate maintenance work orders and send handling instructions, including: When the exception type is personnel violation, a retraining work order is generated and sent to the corresponding personnel's terminal; When the anomaly type is equipment degradation, a maintenance work order is generated and sent to the equipment management terminal; When the anomaly type is spatial obstruction, a scheduling work order is generated and sent to the logistics management terminal; Link and store retraining work orders, maintenance work orders, or scheduling work orders with diagnostic results; The perturbation parameters of the abnormal event operator and the matching score threshold or duration threshold in the preset dispatch conditions are updated based on the judgment accuracy index, including: When the accuracy index of the judgment is lower than the preset accuracy threshold, adjust the time offset parameter, amplitude offset parameter, duration parameter or spatial influence range parameter of the abnormal event operator, and adjust the matching score threshold or duration threshold in the preset distribution conditions. When the accuracy index of the judgment is higher than or equal to the preset accuracy threshold, the current parameters and threshold are retained.

2. The intelligent operation and maintenance method for cleanrooms based on artificial intelligence according to claim 1, characterized in that, Constructing an ideal baseline spacetime tensor includes: Analyze production scheduling data to obtain task timelines; Based on task sequence analysis of personnel permissions and standard operating procedure data, a standard behavior sequence is obtained; Based on task timing, standard behavior sequences, equipment ledger and maintenance record data, and spatial topology configuration data, theoretical baseline power consumption distribution, theoretical pressure difference distribution, and theoretical environmental load distribution are generated. The theoretical baseline power consumption distribution, theoretical pressure difference distribution, and theoretical environmental load distribution are time-aligned and spatially mapped to obtain the ideal baseline spatiotemporal tensor.

3. The intelligent operation and maintenance method for cleanrooms based on artificial intelligence according to claim 1, characterized in that, The abnormal event operator includes time offset parameter, amplitude offset parameter, duration parameter, and spatial influence range parameter; The ideal reference spacetime tensor is parametrically perturbed using anomaly event operators to generate a set of anomalous simulation states, including: The expert rule base is invoked to determine the personnel violation event operator, equipment degradation event operator, and logistics delay event operator; By utilizing the personnel violation event operator to perform personnel behavior perturbations, abnormal personnel simulation states are obtained; The equipment performance is perturbed by using the equipment degradation event operator to obtain the abnormal simulation state of the equipment; By using the logistics congestion event operator to perform spatial blocking perturbation, the abnormal flow field simulation state is obtained; The abnormal simulation states of personnel, equipment, and flow field are summarized to obtain a set of abnormal simulation states.

4. The intelligent operation and maintenance method for cleanrooms based on artificial intelligence according to claim 1, characterized in that, The actual residual matrix is ​​a deviation matrix with time slices as rows and preset monitoring feature dimensions as columns. The preset monitoring feature dimensions include at least one of the following: area power consumption deviation, access control trigger count deviation, door opening duration deviation, particle concentration deviation, pressure difference deviation, and environmental recovery time deviation. Each theoretical residual matrix in the set of theoretical residual matrices has the same dimension as the actual residual matrix; The residuals in the first and second differences are absolute differences, relative deviations, or normalized differences; Before generating diagnostic results, the process also includes: normalizing the set of actual residual matrices and theoretical residual matrices and splicing them according to time windows to construct a unified feature space.

5. An intelligent operation and maintenance system for cleanrooms based on artificial intelligence, characterized in that, It includes a processor and a memory, the memory being coupled to the processor, the memory being used to store computer program code, the computer program code including computer instructions, and the processor calling the computer instructions to cause the system to perform the method as described in any one of claims 1-4.

6. The intelligent operation and maintenance system for cleanrooms based on artificial intelligence according to claim 5, characterized in that, The system includes: The data acquisition module is used to collect business data sources, IoT data sources, and spatial topology configuration data. The benchmark reconstruction module is used to construct an ideal benchmark spatiotemporal tensor; The abnormal disturbance module is used to generate a set of abnormal simulation states; The difference module is used to generate sets of actual residual matrices and theoretical residual matrices; The decision module is used to generate diagnostic results; The work order dispatch module is used to generate maintenance work orders and send processing instructions; The feedback collection module is used to collect execution result data and review result data; The evaluation and update module is used to generate a judgment accuracy index and update the perturbation parameters of the abnormal event operator as well as the matching score threshold or duration threshold in the preset distribution conditions.

Images