Operation and maintenance strategy reasoning method for model generation rule

Abstract

The invention relates to the technical field of rule reasoning, in particular to an operation and maintenance strategy reasoning method for a model generation rule, which comprises the following steps of: finishing value sorting by adopting a Markov process, so that an action field obtains a quantifiable sequence basis in a value evaluation stage; value sequence information and time sequence information form double-sequence constraints in the same action chain, time sequence features are extracted by adopting a time sequence convolutional network, action fields are made to obtain continuous successive marks in the time dimension, the action connection process depends on factor connection and time successive marks at the same time, and the action connection efficiency is improved. A rule number linking process is constructed under the dual conditions of value continuity and time sequence, and cross positioning is performed through an action triggering chain structure and a causal sequence structure element, so that an output rule link forms a consistent basis in the aspects of action value sorting, action causal linking and action time evolution; and the regular structure obtains stability, distinction degree and reusability in a complex operation and maintenance scene.

Description

Technical Field

[0001] This invention relates to the field of rule-based reasoning technology, and in particular to a method for reasoning operational strategies based on model-generated rules. Background Technology

[0002] The field of reasoning technology includes computational methods for knowledge representation, state judgment, and decision generation based on formal rule sets. This field typically consists of three parts: a rule base, a fact base, and a reasoning engine. The rule base records multiple rules represented in the form of conditions to actions. The fact base records the current system state parameters. The reasoning engine triggers rules that meet the conditions from the fact base through forward reasoning, backward reasoning, hybrid reasoning, and conflict resolution strategies, generating multiple decision conclusions. In scenarios such as operation and maintenance control, alarm handling, fault diagnosis, and strategy orchestration, rule-based reasoning technology enables strategy deduction and selection in complex systems using interpretable rule forms.

[0003] A model-generated rule-based operation and maintenance strategy reasoning method is a technical solution that generates multiple operation and maintenance rules through a model and outputs operation and maintenance strategies based on rule reasoning in an operation and maintenance scenario. The method is based on an operation and maintenance object model, an operation and maintenance state model, and an operation and maintenance target model. It automatically generates multiple operation and maintenance rules that conform to a preset rule template structure and loads these multiple operation and maintenance rules into an inference engine for inference. The goal is to stably generate an executable set of operation and maintenance rules from a structured model, and then automatically calculate the combination of operation and maintenance strategies that meet the constraints in complex operation and maintenance scenarios. This method reduces the number of times rules need to be written, reduces the proportion of operation and maintenance strategies that rely on experience, shortens the average time from the occurrence of a fault to the provision of a strategy suggestion, and improves the consistency and reusability of multiple operation and maintenance decisions at the rule level.

[0004] Existing technologies rely on static rule bases to record the mapping between conditions and actions. The lack of a continuous numbering system among condition, state, and action fields leads to unstable triggering order in multi-state alternation scenarios. Furthermore, existing technologies primarily rely on manual rule maintenance, lacking a unified factor system for the condition-to-action structure. This results in a lack of comparable criteria for action fields across different rules, and rule conflicts and overlaps frequently occur in large-scale operations and maintenance scenarios. During the inference process, existing technologies fail to establish a unified time series record for action generation time, action trigger sequence, and action impact range. This causes the inference engine to output strategies that do not conform to the state evolution order when handling alarm escalation chains and resource usage change chains. Moreover, existing technologies rely on static formalized rules, which struggle to synchronously reflect state shifts when dealing with rapid changes in the state of operational objects. This leads to rule response lags when alarm conditions change continuously, making it difficult for strategy outputs to promptly address fault propagation paths in operations and maintenance scenarios. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for reasoning operational strategies based on model generation rules.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for reasoning operation and maintenance strategies based on model generation rules, comprising the following steps: Step 1: Read the three types of data one by one through the real-time status data of the operation and maintenance object, the real-time alarm level data, and the real-time resource occupancy data, and number them in sequence. Then, integrate the numbered content into a status interval arrangement structure according to the reading order to establish the status interval constituent unit. Step 2: Based on the state interval constituent unit, read the action field and sequentially attach the recovery duration factor, risk level factor, and resource change factor. Sort the attached action content according to the Markov process to generate an action value sequence box. Step 3: Based on the action value sequence box, match the state interval number of the condition field item by item and filter the action field. Connect the filtered action fields to the corresponding factors in sequence to form a continuous trigger chain and obtain the action trigger chain structure. Step 4: Based on the action trigger chain structure, generate time points for each action field, extract time features using a temporal convolutional network, compare the relationships before and after comparison, rearrange the compared action fields in chronological order to form a causal connection path, and construct a causal sequence structure element. Step 5: Based on the action trigger chain structure and the causal sequence structure element, locate the action fields of the two structures one by one and determine the corresponding rule number. Connect the numbers in causal order to form a rule link and output the rule link outline.

[0007] As a further embodiment of the present invention, the state interval constituting unit includes a set of state interval numbers corresponding to real-time status data of maintenance objects, real-time alarm level data, and real-time resource occupancy data arranged in numerical order; the action value sequence box includes action fields, recovery duration factors, risk level factors, resource change factors, and action value factors derived from Markov decision processes; the action triggering chain structure includes action fields filtered according to state interval numbers and recovery duration factors, risk level factors, and resource change factors connected to each action field; the causal sequence structure element includes time points, action fields corresponding to time points, and causal connection paths formed by extracting time series features from temporal convolutional networks; the rule link outline includes a rule number sequence arranged in causal order and rule links corresponding to rule numbers.

[0008] As a further aspect of the present invention, the specific steps for establishing the state interval constituting unit are as follows: Based on real-time status data of the operation and maintenance objects, real-time alarm level data, and real-time resource occupancy data, the data is read one by one and the data order is recorded. After writing the number, the content corresponding to the number is stored. The number records are organized and the sequence is kept continuous to generate a number sequence set. Based on the number sequence set, the numbers are extracted one by one and continuous paragraphs are segmented. After the paragraphs are spliced ​​together, the paragraph order is arranged, the paragraph entities are solidified to form interval arrangement records and maintain the continuous structure of the paragraphs, and state interval constituent units are established.

[0009] As a further aspect of the present invention, the specific steps for generating the action value sequence frame are as follows: Based on the state interval constituting unit, the action fields are read one by one and the record positions are established according to the field order. The recovery time factor and the risk level factor are written into the record position. The action value factor from the Markov decision process is written into the record position and the resource change factor is written into the record position. After sorting and writing the content in the order, the action additional set is generated. Based on the action addition set, the additional content is extracted in the order of the records and the difference is read from adjacent records. The difference is written to the corresponding record and the record position is moved according to the difference. After the overall sorting of the records is completed, the sorting is kept continuous to obtain the action difference sequence. Based on the action difference sequence, the sorted content is read one by one and the action field is reorganized according to the sorting order. After merging the additional factor content, the reorganized entries are sorted and reorganized in the order of the marked sequence to form an action arrangement structure and solidify the structure content, thus generating an action value sequence box.

[0010] As a further aspect of the present invention, the Markov decision process includes: based on the existing action value factors and state interval numbers associated with action fields in the action supplement set, reading state transition values ​​one by one from the state transition records corresponding to the state interval numbers; locating reachable states sequentially according to the order of appearance of action fields, and writing the transition values ​​of reachable states into independent transition storage locations; reading the corresponding value quantities from the action value factor set according to the same sequence number position, generating state-action relationship entries by combining the value quantities and the contents of the transition storage locations in a fixed order, and storing the entries in a centralized manner according to the record order as a state-action sequence that can be used for subsequent sorting processing.

[0011] As a further aspect of the present invention, the specific steps for obtaining the action triggering chain structure are as follows: Based on the action value sequence frame, the action fields in the action value sequence frame are extracted sequentially and the condition fields are read synchronously. After extracting the state interval number, the corresponding action field position is located by the number. The position is compared and the action fields that meet the number are selected. The selected content is arranged according to the extraction order to generate an action selection sequence. Based on the action screening sequence, the action screening sequence content is extracted item by item and the recovery time factor is connected in the order of the content. The risk level factor is connected and the resource change factor is connected. After the items are arranged continuously during the connection process, they are organized into a chain structure and the chain record is solidified to obtain the action trigger chain structure.

[0012] As a further aspect of the present invention, the specific steps for obtaining the causal sequence structure element are as follows: Based on the action triggering chain structure, action fields are extracted sequentially and time points are generated before the fields. After writing time markers, the action sequence positions are recorded to form a time-action correspondence list, generating a time point sequence. Based on the time point sequence, the time points are read and time sequence features are extracted by combining a temporal convolutional network. After comparing the time, the sequence is marked with a chronological label. The order of the action fields is adjusted according to the label, and the reordered index is recorded to establish a time-action correspondence table, thereby generating a time-series action set. Based on the time-series action set, action fields and factors are extracted in chronological order. The action fields and factors are then concatenated and written into the causal path unit, followed by appending connection records to obtain the causal sequence structure element.

[0013] As a further aspect of the present invention, the step of extracting time series features using a temporal convolutional network is as follows: Based on the corresponding positions of time markers and action fields recorded in the time point sequence, the time markers are input into the convolution receiver of the temporal convolutional network in the order of appearance, and adjacent time markers are combined into local time segments along a fixed convolution stride after input. The time segments are slid forward sequentially according to the kernel width to generate multiple sets of time segment response values; The response values ​​of each time segment are written to the corresponding intermediate points of the time series processing, and the response values ​​are arranged in a consistent order according to the time point sequence, so that each action field obtains the time series response record corresponding to the time point at the network output.

[0014] As a further aspect of the present invention, the step of extracting action fields and extracting factors in chronological order includes: Based on the existing sorting results within the action set in chronological order, locate each action record in sequence and fully read the action field content stored in each record; The recovery duration factor, risk level factor, and resource change factor corresponding to the action field in the same record are separated from the record structure one by one, while keeping the original position and order of the factors in the record unchanged during the separation process; Each record is labeled with a sequential number according to the time sequence. The action field is then associated with the various factors that have been separated, forming a combined data row that includes the time sequence number, the action field content, and the factor content.

[0015] As a further aspect of the present invention, the specific steps for outputting the rule link outline are as follows: Based on the action trigger chain structure and the causal sequence structure element, the action fields within the two structures are extracted and cross-location points are established with field numbers. The fields of the location points are compared and the comparison positions are recorded. The comparison results are written into the rule number and the number sequence content is organized to generate the rule number sequence. Based on the rule number sequence, each rule number is read and a connection point is established according to the order of the number. The numbers are connected at the connection points and the connection order is recorded. The connection content is pushed into the link structure, the link arrangement is organized and the link record is solidified, and the rule link outline is output.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. In this invention, value ranking is completed by using a Markov process, so that the action field obtains a quantifiable sequence basis in the value assessment stage, and the sequential relationship of actions forms a comparable chain in the value dimension. Moreover, the value sequence information and the time sequence information form a dual sequence constraint in the same action chain. The action screening process depends on both the state interval number and the action value ranking. 2. In this invention, by using a temporal convolutional network to extract time series features, the action field obtains continuous sequential markings in the time dimension, allowing the action records to form a verifiable sequence in causal inference. Moreover, the action connection process depends on both factor connection and temporal sequential markings, so that the rule numbering connection process is completed under the dual conditions of value continuity and temporal order. 3. In this invention, the action trigger chain structure and the causal sequence structure are cross-located to make the output rule chain consistent in terms of action value ranking, action causal connection and action time evolution. Moreover, the rule structure gains stability, distinguishability and reusability in complex operation and maintenance scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Example 1 Please see Figure 1 This invention provides a technical solution: a method for reasoning operation and maintenance strategies based on model generation rules, comprising the following steps: Step 1: Read the three types of data one by one through the real-time status data of the operation and maintenance object, the real-time alarm level data, and the real-time resource occupancy data, and number them in sequence. Then, integrate the numbered content into a status interval arrangement structure according to the reading order to establish the status interval constituent unit. Step 2: Based on the state interval constituent units, read the action field and sequentially attach the recovery duration factor, risk level factor, and resource change factor. Sort the attached action content according to the Markov process to generate the action value sequence box. Step 3: Based on the action value sequence box, match the state interval number of the condition field item by item and filter the action field. Connect the filtered action fields to the corresponding factors in sequence to form a continuous trigger chain and obtain the action trigger chain structure. Step 4: Based on the action trigger chain structure, generate time points for each action field and extract time features by combining a temporal convolutional network. Then compare the relationships before and after, rearrange the compared action fields in chronological order to form a causal connection path, and construct a causal sequence structure element. Step 5: Based on the action trigger chain structure and the causal sequence structure, locate the action fields of the two structures one by one and determine the corresponding rule number. Connect the numbers in causal order to form a rule link and output the rule link outline.

[0020] The state interval constituent unit includes a set of state interval numbers corresponding to real-time status data of operation and maintenance objects, real-time alarm level data, and real-time resource occupancy data arranged in numerical order. The action value sequence box includes action fields, recovery duration factors, risk level factors, resource change factors, and action value factors derived from Markov decision processes. The action trigger chain structure includes action fields filtered by state interval numbers and recovery duration factors, risk level factors, and resource change factors connected to each action field. The causal sequence structure element includes time points, action fields corresponding to time points, and causal connection paths formed by extracting time series features from temporal convolutional networks. The rule link outline includes a rule number sequence arranged in causal order and rule links corresponding to rule numbers.

[0021] The specific steps for establishing the state interval constituent unit are as follows: Based on real-time status data of the operation and maintenance objects, real-time alarm level data, and real-time resource occupancy data, the data is read one by one and the data order is recorded. After writing the number, the content corresponding to the number is stored. The number records are organized and the sequence is kept continuous to generate a number sequence set. Based on the number sequence set, the numbers are extracted one by one and continuous paragraphs are segmented. After the paragraphs are spliced ​​together, the paragraph order is arranged, the paragraph entities are solidified to form interval arrangement records and maintain the continuous structure of the paragraphs, and state intervals are established as units. Based on real-time status data of maintenance objects, real-time alarm level data, and real-time resource occupancy data, the data is read and recorded sequentially, starting from the earliest to the latest time field. The maximum number of records is set to 100,000. The data is then written to the ID field starting from index 1 and incrementing by 1. The threshold values ​​for the real-time status of maintenance objects are set to 0 to 100, the threshold values ​​for real-time alarm levels are set to 1 to 5, and the threshold values ​​for real-time resource occupancy ratios are set to 0 to 100. The ID field is combined with the three types of numerical fields and written into a fixed-length array structure. After traversal, a continuous... The verification method compares the number sequence one by one. The threshold for the difference between adjacent numbers is set to 1. When the difference between adjacent numbers is greater than 1, an empty record is inserted into the array, which only contains the missing number value and uses the status value 0, alarm level 1, and resource usage ratio 0 as placeholder values. After the number verification is completed, the array position is rearranged in order from 1 to the maximum number so that the array index corresponds to the number value. During the rearrangement process, the upper limit threshold of the array length is set to 100000 to prevent the number from going out of bounds. After the rearrangement is completed, the positional relationship between the number field and the three types of numerical fields in the array is used as the reference basis for subsequent processing to generate a number sequence set. Based on the number sequence set, the numbers are extracted one by one and continuous segments are segmented. A sequential traversal is performed, starting from number 1 and scanning forward with a step size of 1. During the scanning process, the difference between the current number and the previous number is calculated. When the difference equals 1, the current number is written to the current segment cache. When the difference is greater than 1, the current segment cache is written as a continuous segment to the segment list, and a new segment cache is opened with the current number. When processing the segment cache, the minimum segment length threshold is set to 5, the maximum segment length threshold is set to 500, and the maximum number of continuous segments is set to 1000. After scanning all numbers, the segment list is sorted in ascending order by the starting number of each segment. During comparison, only the starting number of the current segment is considered. When the starting number of the next segment is greater than the starting number of the next segment, the positions of the two segments are swapped to avoid invalid swaps. After splicing the segments, the segment order is arranged. After sorting, the starting number, ending number, and all numbers in each segment are written into a two-dimensional array to form a three-field structure. The upper limit threshold for the number of rows in the array is set to 1000, and the upper limit threshold for the number of numbers in each row of the array is set to 500. By fixing each row in the two-dimensional array as a continuous number record, the segment entity is solidified to form an interval arrangement record and maintain the continuous structure of the segment. The starting number and ending number of each segment are used as the interval boundary, and the set of numbers in the segment is used as the interval covering number set. The interval arrangement record as a whole is regarded as a set of state intervals, and the state interval constituent units are established.

[0022] The specific steps for generating action value sequence boxes are as follows: Based on the state interval composition unit, the action fields are read one by one and the record positions are established according to the field order. The recovery time factor and the risk level factor are written into the record position. The action value factor from the Markov decision process is written into the record position and the resource change factor is written into the record position. After sorting and writing the content in the order, the action additional set is generated. Based on the action append set, the appended content is extracted in the order of the records and the difference is read from adjacent records. The difference is written to the corresponding record and the record position is moved according to the difference. After the overall sorting of the records is completed, the sorted continuous structure is maintained to obtain the action difference sequence. Based on the action difference sequence, the sorted content is read one by one and the action field is reorganized according to the sorting order. After merging the additional factor content, the reorganized item marker sequence is sorted and the reorganized items are arranged to form an action arrangement structure and solidify the structure content, generating an action value sequence box. Based on the state interval as the unit, the action field is read one by one and the record position is established according to the field order. The action field is traversed sequentially and the order of appearance in the original input is used as the number. The number range is set from 1 to 1000. The recovery time factor, risk level factor, resource change factor and action value factor are written in the order of the number position. The recovery time factor is recorded in minutes and the original value is limited to 1 to 1440. The risk level factor is written directly as an integer from 1 to 5. The resource change factor is recorded as a percentage and the increase or decrease ratio is limited to -100 to 100. The action value factor reads the value from the preset state transition probability table and the instant return table according to the state number corresponding to the action field through the Markov decision process. After the value is updated by the specified number of iterations of 10 rounds and the discount factor of 0.9, the final result is written to the record position. After all the numbered records are completed, all the record positions in the list are rearranged according to the original order of the action fields to generate the action additional set. Based on the action appendage set, appendage content is extracted in record order and differential reading is performed on adjacent records. By calculating the difference item by item, the recovery time factor, risk level factor, resource change factor, and action value factor are written into the difference by subtracting the value of the previous record from the value of the next record. The difference range is limited: recovery time difference is limited to 0 to 1440, risk level difference is limited to 0 to 4, resource change difference is limited to 0 to 200, and action value difference is limited to 0 to 1000. The four differences and their corresponding record numbers are written into the difference structure and sorted in ascending order of absolute difference value. The process uses quick sort and switches to insertion sort when the number of records is less than or equal to 10 to complete the sorting. After sorting, all difference data are rearranged in the sorted record order to obtain the action difference sequence. Based on the action difference sequence, the sorted content is read one by one and the action fields are reorganized according to the sorting order. The action fields of each record are written into the corresponding position of the reorganized action list through sequential merging, and the recovery time factor, risk level factor, resource change factor and action value factor are written simultaneously. When the same action field number appears, the recovery time factor is taken as the average of the two records and rounded down, the risk level factor is taken as the larger value, the resource change factor is taken as the larger absolute value and the sign is retained, and the action value factor is taken as the larger value. After the reorganization is completed, a sequence number is written for each record according to the current position, ranging from 1 to the total number of records. The segment number is continuously increased according to whether the difference between adjacent action value factors is less than 10. All reorganized entries are sorted to form an action arrangement structure and the structure is written into a fixed storage area to generate an action value sequence box.

[0023] The Markov decision process, based on the existing action value factors and state interval numbers associated with action fields in the action appendage set, reads the state transition values ​​from the state transition records corresponding to the state interval numbers one by one. After reading, it locates the reachable states in the order of appearance of the action fields, writes the transition values ​​of the reachable states into independent transition storage locations, and then reads the corresponding value from the action value factor set according to the same sequence number position. It generates state-action relationship entries by combining the value and the contents of the transition storage locations in a fixed order, and stores the entries in the record order as a state-action sequence that can be used for subsequent sorting processing. Markov decision processes, according to the formula: in: Indicates the state Next action The value of operation and maintenance strategy actions determined by the rules generated by the model. This represents the input state within the state interval during the operation and maintenance strategy reasoning process. This field represents the executable action corresponding to this state. This represents the weighting coefficient of the recovery time factor on the value of the action. Indicates the state Execute action The recovery time factor value generated at that time, This represents the weighting coefficient of the risk level factor on the value of the action. Indicates the state Execute action The risk level factor value generated at that time, This represents the weighting coefficient of the impact of resource change factors on the value of actions. Indicates the state Execute action Values ​​of resource change factors over time. This represents the discount factor for the state value after the action is performed. Indicates action In state The subsequent state that may be transitioned to after execution is: Indicates from state Execute action Transition to state The state transition probability, Represents the successor state State value, This represents the set of states consisting of all reachable state intervals; Execution process: The system in the state set Select the current state And determine the action that can be performed. The system calculates the action value function for this state and action combination. The system is based on weighting coefficients. , and Recovery time factor Risk level factor and resource variation factors A linear weighting is applied, and the three factors are included in the action value structure in negative form to quantify the recovery costs, risk impact, and resource changes generated by action execution; a discount factor is also introduced into the system. The value assessment after the state transition is reduced by considering all possible successor states. Calculate state transition probability Value of successor state The product of these factors and their summation form the weighted expectation of the future state value after the action is executed. The system then combines this weighted expectation with the weighted result of the three factors to obtain the action value. .

[0024] The specific steps to obtain the action trigger chain structure are as follows: Based on the action value sequence box, the action fields in the action value sequence box are extracted sequentially and the condition fields are read synchronously. After extracting the state interval number, the corresponding action field position is located by the number. The position is compared and the action fields that meet the number are selected. The selected content is organized according to the extraction order to generate the action selection sequence. Based on the action screening sequence, the content of the action screening sequence is extracted item by item and the recovery time factor is connected in the order of the content. The risk level factor is connected and the resource change factor is connected. After the items are arranged continuously during the connection process, they are organized into a chain structure and the chain record is solidified to obtain the action trigger chain structure. Based on the action value sequence frame, action fields within the sequence frame are extracted sequentially, and condition fields are read synchronously. A sequential traversal is performed, starting from index 1 and progressing to index N with a step size of 1. At each index, the current action field and corresponding condition field are read. The state interval number is extracted from the condition field, with a value range of 1 to 100000. A hash mapping is used to look up a pre-built mapping table, using the state interval number as the key and the corresponding action field's position index within the action value sequence frame as the value. During traversal, the extracted state interval number is accessed to read the corresponding position index from the mapping table and compared with the current traversal position index. When the comparison result is equal, [the action field is selected]. The current action field, along with the status interval number and condition field, are written to the filtering cache array. The maximum length of the filtering cache array is set to 10,000 records. In each record, the action field occupies one string slot, the status interval number occupies one integer slot, and the condition field occupies one string slot. After traversal, sequential writing is used to copy the data to the result array from index 1 to the total number of filtered entries in the original writing order in the filtering cache array. The row index is consistent with the filtering order. During the copying process, a sequence number field is appended to each record. The sequence number range is set to 1 to the total number of filtered entries and written as an integer value. All filtered content is sorted and arranged in the extraction order to generate an action filtering sequence. Based on the action screening sequence, the content of the action screening sequence is extracted item by item, and the recovery duration factor is connected in the order of the content. A chained connection is used, with the sequential number 1 to M in the action screening sequence as the traversal range. The action field index is read from each item, and the corresponding record position in the action value sequence box is accessed. The recovery duration factor value is read from the position. The recovery duration factor value range is set to 1 to 1440 minutes. The value is written into the recovery duration factor slot in the chained structure node, connected to the risk level factor and the resource change factor. Within the same node, the action field, recovery duration factor, risk level factor, and resource change factor are written in a fixed field order: risk level factor value range is set to 1 to 5 (integer), and resource change factor value range is set to -100 to 100% (percentage). The node is then divided... During the matching process, sequential writing is used to record the successor node index for each node in the form of a one-way pointer. When the current node's sequential number is less than M, the successor node index is written with the current number plus 1. When the current node's sequential number is equal to M, the successor node index is written with 0, indicating the end of the chain. After maintaining the continuous arrangement of entries during the connection process, all nodes are sequentially numbered from 1 to M and a complete traversal is performed to verify the continuity of the pointers. Only when the pointer is found to be inconsistent with the sequential number is a single correction made to avoid repeated operations. After the traversal is completed, the contents of all nodes are written to the persistent storage medium in the form of an array. Each node occupies a fixed length of record structure, including an action field (string), a recovery duration factor (integer), a risk level factor (integer), a resource change factor (integer), and a successor node index (integer). After organizing the chain records, an action-triggered chain structure is obtained.

[0025] The specific steps to obtain the causal sequence structure element are as follows: Based on the action trigger chain structure, the action fields are extracted sequentially and time points are generated before the fields. After writing the time markers, the action sequence positions are recorded to form a time-action correspondence list, generating a time point sequence. Based on the time-location sequence, the time-location is read and the time-series features are extracted by combining the temporal convolutional network. After comparing the time, the sequence is marked with a chronological label. The order of the action fields is adjusted according to the label and the reordered index is recorded to establish a time-action correspondence table and generate a time-series action set. Based on the time-series action set, action fields and factors are extracted in time order. The action fields and factors are concatenated and written into the causal path unit, and then connection records are appended to obtain the causal order structure element. Based on the action-triggered chain structure, action fields are extracted sequentially, and time points are generated before the fields. Sequential traversal is used, with the sequence number in the chain ranging from 1 to 1000, and each action field is visited sequentially with a step size of 1. During each visit, a time point number is assigned to the current action field. The time point number is generated starting from 0 milliseconds and incrementing at fixed intervals of 100 milliseconds. The time point value is written into a dedicated time stamp field. At the same time, the action sequence position is recorded in the position field, represented by integers from 1 to 1000. The time stamp field and the action field identifier field are combined and written into the sequence list. Each record includes three items: time point value, action field identifier, and sequence position. The maximum length of the list is 1000 records. After the traversal, the sequence list is checked once according to the sequence position from 1 to the current number of records. Only when a sequence gap is found, the missing position number is added. The arrangement of time point values ​​is not changed. After the sequence check is completed, the sequence list is used as the time and action corresponding records to generate a time point sequence. Based on the time-stamp sequence, time points are read and combined with a temporal convolutional network to extract time-series features. After comparison with the time sequence, chronological markers are added. The time-stamp sequence is then organized into a fixed-length input vector based on the numerical values ​​of the time points, from smallest to largest. The input length is set to 128. If there are fewer than 128 time points, they are padded with zeros. A one-dimensional temporal convolutional network is used. The input layer receives the numerical sequence of time points. The first convolutional layer has 16 kernels, a kernel width of 3, a stride of 1, and an expansion rate of 1. Zero padding is used to maintain the sequence length. The second convolutional layer has 32 kernels, a kernel width of 3, a stride of 1, and an expansion rate of 2. The third convolutional layer has 64 kernels, a kernel width of 3, a stride of 1, and an expansion rate of 1. To achieve a maximum pooling layer of 4, a one-dimensional max pooling layer is added after each convolutional layer. The pooling window width is set to 2 and the stride is set to 2. After three layers of convolution and pooling, a fully connected layer is set at the end to map the output of each position to a one-dimensional scalar to represent the relative order number of the time point in the sequence. After the network forward computation is completed, the output scalars of all time points are sorted in ascending order of value and a sequence marker field is written for each time point, represented by integers from 1 to the number of the current time point. The original action field is rearranged in ascending order of sequence marker and a rearrangement index is written for each rearranged position to record the correspondence between the original sequence number and the new sequence number. The sequence marker, action field, new sequence number, and rearrangement index of the time points are all written into the time and action correspondence table to generate a time-sequence action set. Based on a time-sequential action set, action fields and factors are extracted in chronological order. Sequential traversal is used, with the traversal range from 1 to the current record number as the sequence number of the action fields in the time-sequential action set. Action fields and associated factors, including recovery time factor, risk level factor, resource change factor, and action value factor, are read sequentially from each record. The action fields and four factors within the same record are written into a single causal path unit. The field order is fixed as: time sequence marker, action field, recovery time factor, risk level factor, resource change factor, and action value factor. During the writing process, a unit pointer is added to each causal path unit in an incrementing order. The previous unit pointer records the previous unit's sequence marker value. The previous unit pointer of the first record is written as 0, indicating no predecessor unit. After traversal, all causal path units are arranged in ascending order of sequence marker and written to a contiguous storage area. The total number of units is also saved for later reference. Each row in the contiguous storage area corresponds to a causal path unit. The row content includes six and seven fields: time sequence marker, action field, recovery time factor, risk level factor, resource change factor, action value factor, and previous unit pointer. These are arranged to form a chronologically linked structure of action fields and factors, resulting in a causal sequence structure element.

[0026] Temporal convolutional networks, based on the time markers recorded in the time sequence and their corresponding positions in the action fields, input the time markers one by one into the convolution receiver of the temporal convolutional network in the order of their appearance. After input, adjacent time markers are combined into local time segments along a fixed convolution stride. The time segments are then slid forward sequentially according to the width of the convolution kernel to generate multiple sets of time segment response values. Each time segment response value is written into the corresponding intermediate point of the time sequence processing, and the response values ​​are arranged in a consistent order according to the time sequence, so that each action field obtains a time sequence response record corresponding to the time point at the network output. Temporal convolutional networks, according to the formula: in: Indicates time index The time-order rearranged feature values ​​are calculated by the model generation rules. This refers to the activation function used to introduce nonlinear mappings in an operational strategy reasoning method for model generation rules. This represents the position index on the temporal convolution kernel. This indicates the size of the convolution kernel along the time axis. Indicates the position index of the convolution kernel The weighting coefficient at the location, Indicates time index Location by time point The input vector is formed by concatenating the encodings of its associated action fields. This represents the bias term corresponding to the current convolutional channel in a temporal convolutional network. This represents the weighting coefficient of the time interval feature in the calculation of the time-order rearrangement feature. Indicates time index The normalized time interval characteristics of adjacent time points, This represents the weight coefficient of the action priority feature in the calculation of the temporal reordering feature. Indicates time index The priority scoring features of the action field are determined by the rules generated by the model. This represents the weighting coefficient of the time reliability feature in the calculation of the time order rearrangement feature. Indicates time index Reliability scoring characteristics at specific time points. This represents the weighting coefficient of the time density feature term in the calculation of time-order rearrangement features. Indicates time index The time density characteristics obtained and normalized with this time as the center are statistically derived. Indicates time index The time value at that location; Execution process: The system first reads time points from the time point sequence according to their indices. and time points The input vector is formed by concatenating the encoding of the action field associated with that time point. And thus in time index With time index The time difference is calculated between the time intervals and normalized using the maximum time interval to obtain the time interval feature. The system indexes time according to preset operation and maintenance strategy rules. The execution priority evaluation of the associated action field generates action priority features. The reliability score characteristics of time points are calculated based on the timestamp source calibration method and the consistency of multi-source time records. Meanwhile, indexed by time The time density feature is obtained by counting the number of time points falling within the central time window and dividing by the window length. Subsequently, the system uses a convolution kernel size in the temporal convolutional network. and convolution kernel weight coefficients Regarding the past Input vector at each time step Index by Timeline Perform weighted summation and add bias terms And add a time interval feature term to the above linear combination results. Action priority feature Time reliability characteristics and time density feature terms The weighting coefficient With convolution kernel weights and bias terms The system obtains a mean squared error loss by constructing a monitoring signal with time-series labels and combining it with gradient descent for iterative training. The system then inputs the above linear combination result into the activation function. Calculate the time index Time-order rearranged eigenvalues And rearrange the feature values ​​according to the time order corresponding to all time indices. Sort and reorder the indexes for multiple action fields based on their execution time.

[0027] The action fields and factors are extracted in chronological order. First, based on the existing sorting results within the action set, each action record is located sequentially. The action field content stored in each record is read completely. The recovery time factor, risk level factor, and resource change factor corresponding to the action field in the same record are separated from the record structure one by one. During the separation process, the original position order of the factors in the record is kept unchanged. Each record is labeled with a sequential number according to the time order. The action fields and the various factors separated are established according to the number to form a combined data row containing the time sequence number, action field content, and factor content.

[0028] The specific steps for outputting the rule chain outline are as follows: Based on the action trigger chain structure and the causal sequence structure, the action fields within the two structures are extracted and cross-location points are established using the field numbers. The fields of the location points are compared and the comparison positions are recorded. The comparison results are written into the rule number and the number sequence content is organized to generate the rule number sequence. Based on the rule number sequence, read the rule number one by one and establish the connection point according to the order of the number. Connect the numbers at the connection point and record the connection order. Push the connection content into the link structure, organize the link arrangement and solidify the link record, and output the rule link outline. Based on the action-triggered chain structure and the causal sequence structure, the action field within each of the two structures is traversed from 1 to 1000 in the action-triggered chain structure to read the action field code and value. Simultaneously, a relationship table is established from action field to time sequence number in the causal sequence structure, from 1 to 1000. Each row in the relationship table contains three items: action field code, time sequence number, and internal index. When traversing the action-triggered chain structure, the relationship table is accessed using the action field code as the key. Hash matching is used in a 2048-slot hash table. The slot index is obtained by converting the action field code to an integer and taking the remainder of 2048. When multiple records exist in a slot, the action field codes are compared one by one in ascending order of the internal index. Only records with identical codes are written to the cross-location point. The cross-location point records the action trigger. The chain-structured sequential numbering and the causal sequential structured element time sequence numbering are used. During the cross-location process, a comparison mark value of 1 is written to each successfully matched record, and a comparison mark value of 0 is written to each unmatched record. After the comparison is completed, all records with a comparison mark of 1 are written with rule numbers in the order of appearance, with a value range of 1 to 500. The rule number of the current record is equal to the rule number of the previous record plus 1 and written to the rule number field. After obtaining all rule numbers, the rule number field is sorted by insertion sort in ascending order of rule number. During the sorting process, adjacent records are swapped only when the rule number of the current record is less than the rule number of the previous record. After the sorting is completed, the rule number sequence index field is rewritten according to the sorting result. The sorted rule number field, the corresponding action field code, and the time sequence number are written to a contiguous storage array to generate the rule number sequence. Based on the rule number sequence, rule numbers are read one by one, and connection points are established according to the order of their appearance. For each record in the rule number sequence from 1 to the total number of rules, a sequential traversal is performed. A connection point is created in the connection table for each rule number, recording the current rule number, the predecessor rule number, and the successor rule number. The predecessor rule number is written as the current rule number minus 1 when the current rule number is greater than 1, and as 0 when the current rule number equals 1. The successor rule number is written as the current rule number plus 1 when the current rule number is less than the total number of rules, and as 0 when the current rule number equals the total number of rules. When connecting numbers at the connection points, the system scans sequentially from 1 to the total number of rules. When the difference between the rule numbers of two adjacent records is greater than 1, a virtual rule record is inserted between the two records, assigning a virtual rule to the record. The virtual rule number is equal to the average of the rule numbers of the two preceding and succeeding records, rounded down, and written with the value 1 in a separate marker field. For non-virtual records, the marker field is written with the value 0. After all the numbers are joined, the link table is topologically sorted according to the predecessor rule number and the successor rule number. During the topological sorting process, each round selects the record with the predecessor rule number equal to 0 and which has not been output as the current output record. After output, the predecessor reference of the rule number is deleted from other records until all records are output. In the topological sorting output order, the connection order is written to each record, starting from 1 and increasing in the output order. The rule number, predecessor rule number, successor rule number and connection order of each output record are written sequentially into the link array row index corresponding to the link arrangement order. After the link array is written, the link array is saved as a read-only storage structure, and the rule link outline is output.

[0029] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for reasoning operational strategies based on model generation rules, characterized in that, Includes the following steps: Step 1: Read the three types of data one by one through the real-time status data of the operation and maintenance object, the real-time alarm level data, and the real-time resource occupancy data, and number them in sequence. Then, integrate the numbered content into a status interval arrangement structure according to the reading order to establish the status interval constituent unit. Step 2: Based on the state interval constituent unit, read the action field and sequentially attach the recovery duration factor, risk level factor, and resource change factor. Sort the attached action content according to the Markov process to generate an action value sequence box. Step 3: Based on the action value sequence box, match the state interval number of the condition field item by item and filter the action field. Connect the filtered action fields to the corresponding factors in sequence to form a continuous trigger chain and obtain the action trigger chain structure. Step 4: Based on the action trigger chain structure, generate time points for each action field, extract time features using a temporal convolutional network, compare the relationships before and after comparison, rearrange the compared action fields in chronological order to form a causal connection path, and construct a causal sequence structure element. Step 5: Based on the action trigger chain structure and the causal sequence structure element, locate the action fields of the two structures one by one and determine the corresponding rule number. Connect the numbers in causal order to form a rule link and output the rule link outline.

2. The operation and maintenance strategy reasoning method for model generation rules according to claim 1, characterized in that, The status interval constitutive unit includes a set of status interval numbers corresponding to real-time status data of operation and maintenance objects, real-time alarm level data, and real-time resource occupancy data arranged in numerical order. The action value sequence box includes an action field, a recovery time factor, a risk level factor, a resource change factor, and an action value factor derived from a Markov decision process. The action trigger chain structure includes action fields filtered by state interval numbering, as well as recovery time factor, risk level factor and resource change factor connected to each action field. The causal sequence structure element includes a time point, an action field corresponding to the time point, and a causal connection path formed by extracting time series features from a temporal convolutional network. The rule chain outline includes a sequence of rule numbers arranged in causal order and rule chains corresponding to the rule numbers.

3. The operation and maintenance strategy reasoning method for model generation rules according to claim 1, characterized in that, The specific steps for establishing the state interval constitutive unit are as follows: Based on real-time status data of the operation and maintenance objects, real-time alarm level data, and real-time resource occupancy data, the data is read one by one and the data order is recorded. After writing the number, the content corresponding to the number is stored. The number records are organized and the sequence is kept continuous to generate a number sequence set. Based on the number sequence set, the numbers are extracted one by one and continuous paragraphs are segmented. After the paragraphs are spliced ​​together, the paragraph order is arranged, the paragraph entities are solidified to form interval arrangement records and maintain the continuous structure of the paragraphs, and state interval constituent units are established.

4. The operation and maintenance strategy reasoning method for model generation rules according to claim 1, characterized in that, The specific steps for generating the action value sequence box are as follows: Based on the state interval constituting unit, the action fields are read one by one and the record positions are established according to the field order. The recovery time factor and the risk level factor are written into the record position. The action value factor from the Markov decision process is written into the record position and the resource change factor is written into the record position. After sorting and writing the content in the order, the action additional set is generated. Based on the action addition set, the additional content is extracted in the order of the records and the difference is read from adjacent records. The difference is written to the corresponding record and the record position is moved according to the difference. After the overall sorting of the records is completed, the sorting is kept continuous to obtain the action difference sequence. Based on the action difference sequence, the sorted content is read one by one and the action field is reorganized according to the sorting order. After merging the additional factor content, the reorganized entries are sorted and reorganized in the order of the marked sequence to form an action arrangement structure and solidify the structure content, thus generating an action value sequence box.

5. The operation and maintenance strategy reasoning method for model generation rules according to claim 4, characterized in that, The Markov decision process includes: Based on the existing action value factors and state interval numbers associated with action fields in the action attachment set, read the state transition values ​​from the state transition records corresponding to the state interval numbers one by one. The reachable states are located sequentially according to the order in which the action fields appear, and the transition values ​​of the reachable states are written to independent transition storage locations. Read the corresponding value from the action value factor set according to the same sequence number position, generate state action relationship entries by combining the value and the contents of the transfer storage point in a fixed order, and store the entries in the record order as a state action sequence that can be used for subsequent sorting processing.

6. The operation and maintenance strategy reasoning method for model generation rules according to claim 1, characterized in that, The specific steps to obtain the action trigger chain structure are as follows: Based on the action value sequence frame, the action fields in the action value sequence frame are extracted sequentially and the condition fields are read synchronously. After extracting the state interval number, the corresponding action field position is located by the number. The position is compared and the action fields that meet the number are selected. The selected content is arranged according to the extraction order to generate an action selection sequence. Based on the action screening sequence, the action screening sequence content is extracted item by item and the recovery time factor is connected in the order of the content. The risk level factor is connected and the resource change factor is connected. After the items are arranged continuously during the connection process, they are organized into a chain structure and the chain record is solidified to obtain the action trigger chain structure.

7. The operation and maintenance strategy reasoning method for model generation rules according to claim 1, characterized in that, The specific steps to obtain the causal sequence structure element are as follows: Based on the action triggering chain structure, action fields are extracted sequentially and time points are generated before the fields. After writing time markers, the action sequence positions are recorded to form a time-action correspondence list, generating a time point sequence. Based on the time point sequence, the time points are read and time sequence features are extracted by combining a temporal convolutional network. After comparing the time, the sequence is marked with a chronological label. The order of the action fields is adjusted according to the label, and the reordered index is recorded to establish a time-action correspondence table, thereby generating a time-series action set. Based on the time-series action set, action fields and factors are extracted in chronological order. The action fields and factors are then concatenated and written into the causal path unit, followed by appending connection records to obtain the causal sequence structure element.

8. The operation and maintenance strategy reasoning method for model generation rules according to claim 7, characterized in that, The steps for extracting time series features using a temporal convolutional network are as follows: Based on the corresponding positions of time markers and action fields recorded in the time point sequence, the time markers are input into the convolution receiver of the temporal convolutional network in the order of appearance, and adjacent time markers are combined into local time segments along a fixed convolution stride after input. The time segments are slid forward sequentially according to the kernel width to generate multiple sets of time segment response values; The response values ​​of each time segment are written to the corresponding intermediate points of the time series processing, and the response values ​​are arranged in a consistent order according to the time point sequence, so that each action field obtains the time series response record corresponding to the time point at the network output.

9. The operation and maintenance strategy reasoning method for model generation rules according to claim 7, characterized in that, The steps of extracting action fields and extracting factors in chronological order include: Based on the existing sorting results within the action set in chronological order, locate each action record in sequence and fully read the action field content stored in each record; The recovery duration factor, risk level factor, and resource change factor corresponding to the action field in the same record are separated from the record structure one by one, while keeping the original position and order of the factors in the record unchanged during the separation process; Each record is labeled with a sequential number according to the time sequence. The action field is then associated with the various factors that have been separated, forming a combined data row that includes the time sequence number, the action field content, and the factor content.

10. The operation and maintenance strategy reasoning method for model generation rules according to claim 1, characterized in that, The specific steps for outputting the rule link outline are as follows: Based on the action trigger chain structure and the causal sequence structure element, the action fields within the two structures are extracted and cross-location points are established with field numbers. The fields of the location points are compared and the comparison positions are recorded. The comparison results are written into the rule number and the number sequence content is organized to generate the rule number sequence. Based on the rule number sequence, each rule number is read and a connection point is established according to the order of the number. The numbers are connected at the connection points and the connection order is recorded. The connection content is pushed into the link structure, the link arrangement is organized and the link record is solidified, and the rule link outline is output.

Images