Data analysis method and system applied to industrial think tank assisted decision-making

Abstract

The application provides a data analysis method and system applied to industrial think tank assisted decision-making, relates to the technical field of industrial data analysis, and first acquires an initial industrial operation state record set uploaded on a time sequence collection link in an industrial field, then carries out industrial state thinking disassembly and hierarchical processing on the initial industrial operation state record set, generates an industrial thinking disassembly transmission chain with a hierarchical causal transmission path, inputs the industrial thinking disassembly transmission chain into a pre-constructed industrial thinking chain distillation migration network, obtains an industrial thinking transmission deduction track record, then determines an industrial regulation and control decision element combination according to the industrial thinking transmission deduction track record, and finally generates an industrial think tank assisted regulation and control instruction sequence according to the industrial regulation and control decision element combination, which is used for driving an industrial field regulation and control execution terminal to change an operation parameter configuration state. The application can accurately mine the causal relationship of industrial data, provide effective decision-making, and improve the industrial production efficiency.

Description

Technical Field

[0001] This invention relates to the field of industrial data analysis technology, and more specifically, to a data analysis method and system for industrial think tank-assisted decision-making. Background Technology

[0002] In the field of industrial production, with the continuous improvement of industrial automation, massive amounts of operational status data are generated on the industrial site. These data contain rich information about the operation of industrial equipment and production processes, which are of reference value for the optimization and decision-making of industrial production.

[0003] Currently, data analysis methods in industrial production primarily focus on simple statistics and intuitive visualization of data. For example, they involve plotting curves of equipment operating parameters to observe trends or performing simple classifications of production process states. However, these methods have significant limitations. On the one hand, they struggle to delve into the complex causal relationships hidden behind the data and fail to reveal the intrinsic connections between industrial equipment operating parameters and production process states. On the other hand, existing methods lack systematic modeling and derivation of industrial thinking, failing to provide comprehensive analysis from a holistic perspective of industrial production. This results in decision-making recommendations that often lack specificity and effectiveness. Faced with complex industrial production scenarios and ever-changing market demands, existing technologies struggle to provide effective decision support and cannot meet the demands of industrial production for efficient and intelligent decision-making. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a data analysis method for industrial think tank-assisted decision-making, the method comprising:

[0005] Acquire the initial industrial operation status record set uploaded by the industrial site time-series acquisition link. The initial industrial operation status record set includes industrial equipment operation parameter snapshot units and production process status snapshot units arranged in the order of acquisition time.

[0006] The initial industrial operation state record set is subjected to industrial state thinking decomposition and hierarchical processing. The state offset information in the industrial equipment operation parameter snapshot unit and the production process state snapshot unit is matched with the preset industrial thinking decomposition template for causal transmission path matching to generate an industrial thinking decomposition transmission chain with a hierarchical causal transmission path. The industrial thinking decomposition transmission chain contains multiple industrial thinking atomic state nodes connected in the order of thinking deduction.

[0007] The industrial thinking decomposition transmission chain is input into a pre-constructed industrial thinking chain distillation migration network. The transmission strength between nodes is calculated node by node along the thinking deduction order of the industrial thinking atomic state nodes, and the intermediate state description of the transmission deduction is generated to obtain the industrial thinking transmission deduction trajectory record. The industrial thinking transmission deduction trajectory record includes the intermediate state description of each step of thinking transmission and the distribution record of thinking transmission strength.

[0008] Based on the description of intermediate states of thought transmission and the distribution of thought transmission intensity in the industrial thought transmission trajectory record, the combination of industrial control decision elements for the target industrial regulation object is determined. The combination of industrial control decision elements includes a control tendency descriptor and a control receptor unit identifier.

[0009] An industrial think tank-assisted control instruction sequence is generated based on the control tendency descriptor and the control receptor unit identifier. The industrial think tank-assisted control instruction sequence is used to drive the industrial field control execution terminal to change the configuration status of operating parameters.

[0010] Furthermore, embodiments of the present invention also provide a data analysis system for industrial think tank-assisted decision-making, comprising:

[0011] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned data analysis method for industrial think tank-assisted decision-making by executing the machine-executable instructions.

[0012] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions stored in a computer-readable storage medium, a processor of a data analysis system for industrial think tank-assisted decision-making reads the machine-executable instructions from the computer-readable storage medium, the processor executes the machine-executable instructions, causing the data analysis system for industrial think tank-assisted decision-making to perform the aforementioned data analysis method for industrial think tank-assisted decision-making.

[0013] Based on the above, by acquiring the initial industrial operation status record set uploaded from the industrial site time-series acquisition link, the initial industrial operation status record set is subjected to industrial state thinking decomposition and hierarchical processing. The state offset information is matched with the preset industrial thinking decomposition template for causal transmission path matching, generating an industrial thinking decomposition transmission chain with a hierarchical causal transmission path. This can reveal the complex causal relationship between industrial equipment operating parameters and production process status, constructing a complete industrial thinking logic framework. The industrial thinking decomposition transmission chain is input into a pre-constructed industrial thinking chain distillation migration network. The transmission strength between nodes is calculated node by node along the thinking deduction sequence of the industrial thinking atomic state nodes, and the intermediate state description of the transmission deduction is generated. This yields an industrial thinking transmission deduction trajectory record containing the intermediate state description of each step of the thinking transmission and the distribution record of the thinking transmission strength. This achieves in-depth deduction and quantitative analysis of industrial thinking, enabling an accurate understanding of the thinking transmission law in the industrial production process. Based on the industrial thinking transmission deduction trajectory record, the combination of industrial control decision elements for the target industrial regulation object is determined. Based on this combination of industrial control decision elements, an industrial think tank-assisted control instruction sequence is generated, effectively improving the decision-making efficiency and accuracy of industrial production, reducing production costs, and improving the overall efficiency of industrial production. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the execution flow of the data analysis method for industrial think tank decision support provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of exemplary hardware and software components of a data analysis system for industrial think tank decision support provided in an embodiment of the present invention. Detailed Implementation

[0016] Figure 1 This is a flowchart illustrating a data analysis method for industrial think tank decision support provided by an embodiment of the present invention, which will be described in detail below.

[0017] Step S110: Obtain the initial industrial operation status record set uploaded by the industrial site time-series acquisition link. The initial industrial operation status record set includes industrial equipment operation parameter snapshot units and production process status snapshot units arranged in the order of acquisition time.

[0018] In this embodiment, the industrial think tank-assisted decision-making system establishes a data communication link with the programmable logic controller and distributed control system through an industrial fieldbus network. This data communication link employs a time synchronization protocol to ensure clock consistency across all data acquisition nodes. The data structure of the initial industrial operating status record set is defined as follows: Let the acquisition time sequence be... Where M is the total number of acquisition times, and the time interval between adjacent acquisition times is ΔT. At each acquisition time tᵢ, the industrial equipment operating parameter snapshot unit E(tᵢ) is a real-valued vector of dimension A. ,in Let P(tᵢ) represent the value of the monitored parameter of the a-th industrial equipment at time tᵢ. The production process status snapshot unit P(tᵢ) is a real-valued vector of dimension B. , where pb(tᵢ) represents the value of the b-th production process state parameter at time tᵢ. All and according to Arranged in ascending order, they form the initial industrial operation status record set. .

[0019] Step S120: Perform industrial state thinking decomposition and hierarchical processing on the initial industrial operation state record set, match the state offset information in the industrial equipment operation parameter snapshot unit and the production process state snapshot unit with the preset industrial thinking decomposition template for causal transmission path, and generate an industrial thinking decomposition transmission chain with a hierarchical causal transmission path. The industrial thinking decomposition transmission chain contains multiple industrial thinking atomic state nodes connected in the order of thinking deduction.

[0020] Step S121: Extract the parameter offset direction indicator of each monitored parameter in each industrial equipment operation parameter snapshot unit in the initial industrial operation status record set, and combine all parameter offset direction indicators of the same industrial equipment operation parameter snapshot unit into a device parameter offset tendency direction vector; and extract the process stagnation degree indicator of each process node in each production process status snapshot unit in the initial industrial operation status record set, and combine all process stagnation degree indicators of the same production process status snapshot unit into a process stagnation tendency direction vector.

[0021] For industrial equipment operating parameter snapshot unit A baseline operating point value is pre-set for each monitoring parameter. and a dead zone threshold Parameter offset The calculation logic is as follows Parameter offset direction indicator The determination rule is: if Greater than ,but ;like Less than ,but ;otherwise Iterate through a=1 to A to obtain A parameter offset direction indicators, and combine them according to the index order of a to form the device parameter offset tendency direction vector. .

[0022] For the production process status snapshot unit A normal operating range [lowb, highb] is pre-defined for each process status parameter. Process stagnation level indicator. The calculation logic is as follows: If If it is greater than highb, then ;like If it is less than lowb, then ;otherwise Iterate through b=1 to B to obtain B process stagnation indicators, and combine them according to the index order of b to form a process stagnation tendency vector. .

[0023] Step S122: Calculate the first direction matching degree sequence between the device parameter offset tendency direction vector and each preset industrial thinking disassembly template, and simultaneously calculate the second direction matching degree sequence between the process stagnation tendency direction vector and each preset industrial thinking disassembly template. Normalize the matching degree values ​​in the first and second direction matching degree sequences, and perform weighted fusion processing on each matching degree value in the normalized first direction matching degree sequence and the matching degree value in the normalized second direction matching degree sequence corresponding to the same order to generate a comprehensive matching degree ranking queue. Select the preset industrial thinking disassembly template corresponding to the first-ranked comprehensive matching degree from the comprehensive matching degree ranking queue as the target industrial thinking disassembly template.

[0024] In this embodiment, K industrial thinking disassembly templates are preset, denoted as TEMPLATEk, k=1,2,...,K. Each TEMPLATEk contains a device expected offset direction vector Weqk=[w1k, w2k,..., wAk] of length A, where... And a process expected blocking direction vector Wprock=[u1k, u2k, ..., uBk] of length B, where ubk∈[0,1].

[0025] For the current time tᵢ and Calculate the first-direction matching degree matcheq,k and the second-direction matching degree matchproc,k with each TEMPLATEk. The calculation logic for matcheq,k is as follows: matcheq, ,in This is an indicator function; it takes a value of 1 when the two are equal, and 0 otherwise. The calculation logic for `matchproc,k` is as follows: `matchproc,` .

[0026] Normalize matcheq,k and matchproc,k to obtain normMatcheq,k = matcheq,k / (Σmatcheq,j) and normMatchproc,k = matchproc,k / (Σmatchproc,j). Set a fusion weight coefficient α ∈ [0,1] and calculate the overall matching degree totalMatchk = α × normMatcheq,k + (1-α) × normMatchproc,k. Sort all K overall matching degrees in descending order to obtain an overall matching degree sorting queue. Select the template index corresponding to totalMatchmax at the top of the queue. Then the target industry thinking decomposition template .

[0027] Step S123: Compare the original observation values ​​of each monitoring parameter contained in the industrial equipment operation parameter snapshot unit with the preset initial offset threshold range in the target industrial thinking decomposition template, and determine the hierarchical position mark of each monitoring parameter in the thinking decomposition starting level based on the comparison results.

[0028] The TARGET TEMPLATE template for industrial thinking decomposition includes a hierarchical mapping table. This table defines the correspondence between the numerical range of each monitoring parameter 'a' and the decomposition level. Specifically, for monitoring parameter 'a', L levels are preset, and each level l (l=1, 2, ..., L) corresponds to a numerical range [lowa, l, higha, l]. Compare the intervals in this group step by step to find those that satisfy lowa. The smallest level index l (i.e., the topmost matching level) of l is used as the layer order marker layera for the monitoring parameter a. If If it does not fall into any preset range, its hierarchical position is set to L (deepest level).

[0029] Step S124: Compare the original state description values ​​of each process node contained in the production process state snapshot unit with the preset initial process threshold range in the target industrial thinking decomposition template, and determine the hierarchical position mark of each process node in the thinking decomposition starting level based on the comparison results.

[0030] Similarly, the TARGET TEMPLATE template for decomposing industrial thinking also predefines L numerical ranges [lowb, l, highb, l] corresponding to each process state parameter b. Compare the intervals in this group step by step to find those that satisfy lowb. The smallest level index l of l is used as the hierarchical position marker layerb of process node b.

[0031] Step S125: For all monitoring parameter offset direction indicators and all process obstruction degree indicators belonging to the same thinking disassembly level, perform offset polarity aggregation statistical processing, extract the mode direction of the offset direction indicators within the thinking disassembly level as the offset direction aggregation representation, and normalize the offset degree value of each type of monitoring parameter within the thinking disassembly level, and then perform a weighted average of the normalized offset degree values ​​of all monitoring parameters to obtain the comprehensive offset magnitude aggregation representation of the thinking disassembly level, wherein the weight is determined according to the preset importance coefficient of each monitoring parameter in the preset industrial thinking disassembly template.

[0032] For each thought decomposition level l (l=1, 2, ..., L), collect the monitoring parameters a for all belonging level sequence markers layera equal to l, and the process nodes b for all belonging level sequence markers layerb equal to l. First, for these monitoring parameters, their offset direction indicators... Create a set Sdir,l. Calculate the mode of this set, which is the direction value that appears most frequently, and use it as the aggregated representation of the offset direction for this level, aggDirl. If the mode is not unique, prioritize +1 or -1, and then 0.

[0033] Secondly, for these monitoring parameters, their offset values ,in To monitor the range of parameter 'a', the calculation normalizes the offset to the [0, 1] interval. For these process nodes, the degree of obstruction is... The values ​​are already in the range [0, 1]. These two types of values ​​are merged into a single offset sequence. The TARGET TEMPLATE template for industrial thinking pre-sets importance coefficients for each monitoring parameter a and process node b. And impb. The calculation logic for the aggregated offset-level representation of aggMagl is as follows: , aggMagl = numerator / denominator.

[0034] Step S126: The offset direction aggregation representation and the offset magnitude aggregation representation are jointly constructed into an industrial thinking atomic state node corresponding to the thinking decomposition level. The industrial thinking atomic state node simultaneously contains a hierarchical position mark, an offset direction aggregation representation, and an offset magnitude aggregation representation.

[0035] For each thinking decomposition level l, an industrial thinking atomic state node NODEl is created. Said industrial thinking atomic state node is a triplet data structure in the form of NODEl=(layerId=l, direction=aggDirl, magnitude=aggMagl). Wherein layerId is a level order mark, direction is an aggregated representation of offset direction, and magnitude is an aggregated representation of offset magnitude. All L nodes form a node list [NODE1, NODE2, ..., NODEL].

[0036] Step S127: according to the pre-established inter-level causal conduction dependency paths in said target industrial thinking decomposition template, establish directed thinking conduction connecting edges between industrial thinking atomic state nodes with different level order marks, wherein the direction of said directed thinking conduction connecting edge points from the industrial thinking atomic state node with a preceding level order mark to the industrial thinking atomic state node with a succeeding level order mark, and combine all industrial thinking atomic state nodes and all directed thinking conduction connecting edges to form said industrial thinking decomposition conduction chain with hierarchical causal conduction paths.

[0037] The target industrial thinking decomposition template TARGET TEMPLATE contains an L×L causal conduction adjacency matrix C, wherein Cl1, l2=1 indicates that there exists a causal conduction dependency relationship from level l1 to level l2, and l1<l2. According to said adjacency matrix, for each pair of level combinations satisfying Cl1, l2=1, a directed thinking conduction connecting edge EDGEl1→l2 pointing from NODEl1 to NODEl2 is established. All NODEl and EDGEl1→l2 together form a directed acyclic graph structure, that is the industrial thinking decomposition conduction chain CHAINTHINK. In said industrial thinking decomposition conduction chain, nodes are arranged in the direction of increasing level order marks, and edges represent the direction of thinking conduction.

[0038] Step S130: input said industrial thinking decomposition conduction chain into a pre-constructed industrial thinking chain distillation transfer network, calculate inter-node conduction intensity node by node along the thinking derivation order of said industrial thinking atomic state nodes and generate conduction deduction intermediate state descriptions, so as to obtain an industrial thinking conduction deduction track record, wherein said industrial thinking conduction deduction track record comprises each step of thinking conduction intermediate state description and a thinking conduction intensity distribution record.

[0039] Step S131: Perform node vectorization processing on all industrial thinking atomic state nodes contained in the hierarchical causal transmission chain of industrial thinking decomposition. Map the offset direction aggregation representation and offset magnitude aggregation representation contained in each industrial thinking atomic state node together into a fixed-length industrial thinking node embedding vector to obtain a fixed-length industrial thinking node embedding vector sequence that corresponds one-to-one with all industrial thinking atomic state nodes. Insert a pre-trained industrial thinking start embedding vector at the beginning of the fixed-length industrial thinking node embedding vector sequence and insert a pre-trained industrial thinking end embedding vector at the end of the fixed-length industrial thinking node embedding vector sequence to generate a complete industrial thinking transmission embedding vector sequence.

[0040] For each node NODE1 in the CHAINTHINK deconstruction chain, its direction field aggDirl ∈ {-1, 0, +1} and magnitude field aggMagl ∈ [0, 1]. First, aggDirl is one-hot encoded into a 3D vector Vdir, l = [I(aggDirl=-1), I(aggDirl=0), I(aggDirl=+1)]. Then, aggMagl is expanded into a 1D vector Vmag, l = [aggMagl]. Vdir, l and Vmag, l are concatenated to obtain a 4D original feature vector Vraw, l. Then, Vraw, l is mapped to a Dmodel-dimensional industrial thinking node embedding vector Hnode, l = Wemb × Vraw, l through a trainable linear mapping matrix Wemb (size Dmodel × 4), where Dmodel is the preset embedding dimension. Perform the above operations on all L nodes to obtain the node embedding vector sequence [Hnode, 1, Hnode, 2, ..., Hnode, L]. Insert a pre-trained industrial thinking start embedding vector Hstart at the beginning of this sequence, with dimension Dmodel, representing the starting point of the thought process. Insert a pre-trained industrial thinking end embedding vector Hend at the end of this sequence, with dimension Dmodel, representing the ending point of the thought process. Finally, obtain the complete industrial thinking transmission embedding vector sequence Hseq = [Hstart, Hnode, 1, Hnode, 2, ..., Hnode, L, Hend], with a total length of L+2.

[0041] Step S132: Add a position encoding vector corresponding to the position of each embedding vector in the complete industrial thinking transmission embedding vector sequence to obtain an industrial thinking transmission embedding vector sequence carrying position encoding. The numerical distribution of the position encoding vector is generated according to a preset sine-cosine alternation function.

[0042] For the pos-th position in the complete industrial thinking transmission embedding vector sequence (pos is counted from 0 to L+1), a Dmodel-dimensional position encoding vector PE(pos) is generated. The numerical calculation logic for the 2i-th dimension of PE(pos) is sin(pos / (100002i / Dmodel)), and the numerical calculation logic for the 2i+1-th dimension is cos(pos / (100002i / Dmodel)), where i=0,1,...,(Dmodel / 2)-1. The position encoding vector is added element-wise to the embedding vector to obtain the industrial thinking transmission embedding vector sequence HposSeq carrying the position encoding, where HposSeq[pos]=Hseq[pos]+PE(pos).

[0043] Step S133: The sequence of industrial thinking transmission embedding vectors carrying position codes is sent to the first transmission deduction step of the industrial thinking chain distillation transfer network. In the first transmission deduction step, for each industrial thinking node embedding vector in the sequence, the vector dot product similarity between the industrial thinking node embedding vector and all industrial thinking node embedding vectors in the sequence that are positioned before and after it is calculated. The calculated vector dot product similarity set is numerically scaled using a preset normalization exponential function to obtain the attention weight distribution value that corresponds one-to-one with all industrial thinking node embedding vectors in the sequence that are positioned before and after it. The attention weight distribution value is then used to perform a weighted summation on all industrial thinking node embedding vectors in the sequence that are positioned before and after it to generate the updated embedding vector after the first transmission deduction step. After traversing all industrial thinking node embedding vectors, the updated embedding vector sequence of the first transmission deduction step is obtained.

[0044] The Industrial Mind Chain Distillation Transfer Network is built on a Transformer decoder architecture and contains Tlayer identical inference layers. Taking the first inference layer as an example, its input is HposSeq. For the pos-th position in the output sequence, its query vector Qpos = HposSeq[pos] × WQ is calculated first, where WQ is a learnable query weight matrix with size Dmodel × Dmodel. For all position indices j satisfying 0 ≤ j ≤ pos, the key vector Kj = HposSeq[j] × WK is calculated, where WK is a learnable key weight matrix with size Dmodel × Dmodel. The dot product similarity scorepos,j = (Qpos·Kj) / √Dmodel is calculated between Qpos and each Kj. All scorepos,j (j = 0, ..., pos) are combined into a vector Spos. The attention weight distribution Apos = softmax(Spos) is obtained by numerically scaling Spos using a normalized exponential function, where Apos[j] = exp(scorepos,j) / (Σexp(scorepos,m)). Then, for all position indices j satisfying 0≤j≤pos, calculate the value vector Vj=HposSeq[j]×WV, where WV is the learnable value weight matrix with size Dmodel×Dmodel. Update the output vector for that position with H'pos=ΣApos[j]×Vj. Perform the above calculation sequentially for pos=0 to L+1 to obtain the output sequence H'seq=[H'0,H'1,...,H'L+1].

[0045] Step S134: The first transmission deduction step update embedding vector sequence is used as the input of the second transmission deduction step. The vector dot product similarity calculation, attention weight distribution value calculation and weighted summation are repeatedly performed to generate the second transmission deduction step update embedding vector sequence. This process is repeated until the preset total number of transmission deduction steps is completed to obtain the final transmission deduction update embedding vector sequence. The starting position update embedding vector corresponding to the industrial thinking starting embedding vector is extracted from it. The starting position update embedding vector is input into the transmission intermediate state description generator of the industrial thinking chain distillation transfer network. According to the input starting position update embedding vector, the corresponding word is retrieved one by one in the preset industrial thinking transmission description vocabulary and output. The retrieved word is concatenated in order to form the thinking transmission intermediate state description corresponding to each thinking transmission step. At the same time as outputting each thinking transmission intermediate state description, the transmission confidence scalar value corresponding to the thinking transmission intermediate state description is calculated.

[0046] The output sequence H'seq of the first propagation layer is used as the input of the second propagation layer, and the above operation is repeated. This process continues until the Tlayer propagation layer is completed. The last propagation layer outputs the final propagation update embedding vector sequence HfinalSeq=[Hfinal, 0, Hfinal, 1, ..., Hfinal, L+1]. The update embedding vector Hfinal, 0 at the starting position (pos=0) is extracted and input into the propagation intermediate state description generator. This propagation intermediate state description generator consists of an autoregressive language model head, containing a linear transformation layer and a flexible maximum transfer function layer. The linear transformation layer maps Hfinal, 0 (dimension Dmodel) to a logits vector with a dimension equal to the preset vocabulary size V. The flexible maximum transfer function converts this logits vector into a word probability distribution. The word with the highest probability is sampled or selected from this word probability distribution as the first output word. Then, the embedding vector of this word is combined with Hfinal, 0 to continue predicting the next word until the end symbol is predicted. The word sequence output during the entire prediction process constitutes the thought propagation intermediate state description text TEXTdesc. When generating each lexical unit, its probability value is accumulated or taken as a geometric mean to obtain the scalar value of the transmission confidence corresponding to the descriptive text, confDesc.

[0047] Step S135: The intermediate state description and confidence scalar value of each thought transmission step are associated and stored in the order of thought transmission steps to generate the industrial thought transmission deduction trajectory record. Each record in the industrial thought transmission deduction trajectory record contains the thought transmission step number, the intermediate state description text, and the confidence scalar value.

[0048] The intermediate state description text TEXTdesc generated in the above process is used as the output of the current thought transmission deduction step. Steps S131 to S134 are repeated to iteratively deduce each node in the industrial thought decomposition transmission chain. After each deduction step is completed, a triple record (stepIdx, TEXTdesc, confDesc) is generated, where stepIdx is the sequence number of the current thought transmission step. All triples of all steps are stored in order of stepIdx to form the industrial thought transmission deduction trajectory record TRAJrecord.

[0049] Step S140: Determine the combination of industrial control decision elements for the target industrial regulation object based on the description of the intermediate state of thought transmission and the distribution record of thought transmission intensity in the industrial thought transmission deduction trajectory record. The combination of industrial control decision elements includes a control tendency descriptor and a control receptor unit identifier.

[0050] Step S141: Extract the intermediate state description text and the transmission confidence scalar value of each thought transmission step sequentially from the industrial thought transmission trajectory record. Associate the intermediate state description text and the transmission confidence scalar value of the same thought transmission step number to form a single-step transmission analysis unit. Perform regulatory semantic component decomposition processing on the intermediate state description text in each single-step transmission analysis unit to obtain the regulatory action semantic fragment and regulatory receptor semantic fragment contained in the intermediate state description text. Match and transform the regulatory action semantic fragment with the preset regulatory action semantic mapping rule set to generate the regulatory tendency pointing descriptor corresponding to the single-step transmission analysis unit. The regulatory tendency pointing descriptor is used to characterize the industrial regulation operation direction type.

[0051] For each item (stepIdx, TEXTdesc, confDesc) in TRAJrecord, TEXTdesc is first parsed using a rule-based tokenizer or named entity recognition model. This parsing process identifies verb phrases in the text as the regulatory action semantic fragment actPhrase and noun phrases as the regulatory receptor semantic fragment tgtPhrase. The pre-defined set of regulatory action semantic mapping rules is a key-value dictionary, where each key is a verb phrase pattern (e.g., "increase", "decrease", "on", "off") and each value is a corresponding regulatory tendency descriptor (e.g., INCREASE, DECREASE, TURNON, TURNOFF). Matching actPhrase with this dictionary yields the regulatory tendency descriptor actionDesc. This regulatory tendency descriptor represents the directional type of the industrial regulatory operation.

[0052] Step S142: Match and convert the semantic fragment of the regulatory receptor with the preset set of industrial entity identifier mapping rules to generate the regulatory receptor unit identifier corresponding to the single-step transmission parsing unit. The regulatory receptor unit identifier is consistent with the unique identifier code of the source equipment of the industrial equipment operation parameter snapshot unit or the source process node of the production process status snapshot unit.

[0053] The pre-defined set of industrial entity identifier mapping rules is also a key-value dictionary. Each key is a name pattern for a device or process node (e.g., "reactor R001", "tower T003"), and each value is a unique identifier code (e.g., EQP R001, COL T003). Matching tgtPhrase with this dictionary yields the regulatory receptor unit identifier targetId. This regulatory receptor unit identifier is consistent with the device identifier code used in the initial industrial operating status record set.

[0054] Step S143: Use the conduction confidence scalar value of the single-step conduction analysis unit as the weighting factor of the corresponding regulatory tendency pointing descriptor and regulatory receptor unit identifier, and combine the regulatory tendency pointing descriptor, regulatory receptor unit identifier and weighting factor into single-step regulation candidate items.

[0055] For actionDesc and targetId obtained from stepIdx in steps S141 and S142, a single-step control candidate entry candEntry=(targetId, actionDesc, weight=confDesc) is constructed by combining the propagation confidence scalar value confDesc of this step.

[0056] Step S144: In the single-step regulation candidate item set, for items with the same regulatory receptor unit identifier, the cumulative score of each regulatory tendency pointing descriptor is calculated by weighting the weight factor. The regulatory tendency pointing descriptor with the highest cumulative score is paired with the regulatory receptor unit identifier as the target regulation decision element. All target regulation decision elements are collected to form the industrial regulation decision element combination for the target industrial regulation object.

[0057] Iterate through all candEntries generated in the TRAJrecord, grouping all entries with the same targetId into one group. For each group, calculate the weighted cumulative score of each actionDesc: score(action) = Σ (weight of all entries in the same group). Select the actionDesc with the highest score as the final regulatory tendency descriptor finalAction corresponding to that targetId. Treat (targetId, finalAction) as a target regulatory decision element decElem. The set of all decElems constitutes the industrial regulatory decision element combination DECset.

[0058] Step S150: Based on the regulation tendency descriptor and the regulation receptor unit identifier, an industrial think tank-assisted regulation instruction sequence is generated. This industrial think tank-assisted regulation instruction sequence is used to drive the industrial field regulation execution terminal to change the configuration status of operating parameters.

[0059] Step S151: Read the target control decision elements one by one from the combination of industrial control decision elements, obtain the control tendency descriptor and control receptor unit identifier contained in each target control decision element, and retrieve the corresponding set of controllable parameter names and the current operating setting value of each controllable parameter name from the preset industrial control parameter configuration library based on the control receptor unit identifier.

[0060] For each decElem=(targetId, finalAction) in DECset, its targetId is used as the key to retrieve the parameter from the preset industrial control parameter configuration library paramDB. paramDB is a key-value database, where each targetId corresponds to a list of controllable parameters: paramList(targetId)=[(paramName1, currentVal1), (paramName2, currentVal2), ...], where paramName is the name of the controllable parameter, and currentVal is the current setting value of the parameter.

[0061] Step S152: For each adjustable parameter name, determine the target adjustment operation to be performed on the current operating setting value of the adjustable parameter name based on the adjustment direction indicated by the adjustment tendency descriptor, and obtain the target operating setting value of the adjustable parameter name.

[0062] For each adjustable parameter (paramName, currentVal) in paramList, the adjustment logic is determined based on the type of finalAction. If finalAction is INCREASE, the target running setpoint targetVal = currentVal × (1 + stepRatio), where stepRatio is the preset single-step adjustment step size ratio. If finalAction is DECREASE, targetVal = currentVal × (1 - stepRatio). If finalAction is TURNON, targetVal = 1 (indicating the on state). If finalAction is TURNOFF, targetVal = 0 (indicating the off state). For parameters that require continuous adjustment, it is also checked whether targetVal exceeds the preset safe operating range [minVal, maxVal]. If it does, it is clamped to the boundary value.

[0063] Step S153: Combine the regulatory receptor unit identifier, the name of each adjustable parameter, and its target operating setting value into a regulatory instruction content entry.

[0064] (targetId, paramName, targetVal) is used as the basic unit of a controllable instruction content entry (cmdContent). If the same targetId corresponds to multiple controllable parameters, all (paramName, targetVal) pairs are combined with targetId to form a complete controllable instruction content entry.

[0065] Step S154: Obtain the terminal addressing identifier of the target control execution terminal that is pre-bound to the control receptor unit identifier; encapsulate the control instruction content item with the terminal addressing identifier to generate an industrial control instruction unit; arrange all industrial control instruction units in the original order of each target control decision element in the industrial control decision element combination to form an industrial think tank-assisted control instruction sequence; and convert it into an instruction data block in a preset industrial control communication protocol format for sending to the target control execution terminal.

[0066] In the industrial field device registry, each targetId is pre-bound to the network address or bus address of a target control execution terminal, denoted as addrTarget. The cmdContent generated in step S153 is encapsulated with this addrTarget to form an industrial control instruction unit cmdUnit conforming to an industrial communication protocol (such as OPCUA or ModbusTCP). The above operation is performed on all decElems in the DECset in their original order to obtain a cmdUnit sequence. This sequence is converted into a continuous instruction data block in a preset industrial control communication protocol format and sent to the corresponding target control execution terminal via the industrial field network to drive it to change its operating parameter configuration status.

[0067] Step S210: Perform the industrial mind chain distillation pre-training process to obtain the network parameter configuration information of the industrial mind chain distillation migration network.

[0068] Before inputting the decomposed industrial thinking transmission chain into the pre-constructed industrial thinking chain distillation transfer network, the network needs to be pre-trained. In this embodiment, the industrial thinking chain distillation transfer network adopts a Transformer decoder architecture, containing 12 identical transmission inference layers. Each transmission inference layer contains a masked multi-head self-attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by layer normalization and residual connections. The multi-head self-attention sublayer contains 8 attention heads, and the hidden layer dimension of the feedforward neural network sublayer is 2048, with an embedding dimension Dmodel of 512. The pre-training process employs a knowledge distillation strategy, using a large-scale language model of the industrial domain with a larger number of parameters as the teacher model.

[0069] Step S211: Obtain the sample thought transmission trajectory record output by the large-scale language model in the industrial field for the sample industrial operation status record set. The sample thought transmission trajectory record includes the sample thought transmission step sequence and the sample thought transmission intermediate state description text corresponding to each sample thought transmission step.

[0070] A sample dataset of 100,000 historical industrial operation status records and their corresponding expert control operation records was collected. Each set of sample industrial operation status records was input into a pre-trained large-scale language model for the industrial domain. This teacher model outputs a corresponding sample thought transmission trajectory record. This sample thought transmission trajectory record is a structured sequence, denoted as TRAJteacher, containing P sample thought transmission steps. Each sample thought transmission step stepp contains a sample thought transmission intermediate state description text TEXTteacher,p.

[0071] Step S212: Each step of the sample thought transmission trajectory record is marked with a step number, and each sample thought transmission intermediate state description text is assigned a unique sequential number in the sample thought transmission step sequence to form a sample thought transmission supervision signal set.

[0072] For each step p in TRAJteacher, its position index p in the sequence is used as its unique sequence number. (p, TEXTteacher, p) is treated as a supervisory signal unit. The set of all P supervisory signal units constitutes the sample thought transmission supervisory signal set SUPset.

[0073] Step S213: Using the same processing method as the initial industrial operation state record set for industrial state thinking decomposition and hierarchical processing, the sample industrial operation state record set is processed to generate a sample industrial thinking decomposition transmission chain set corresponding to the sample industrial operation state record set.

[0074] For each set of data in the sample industrial operation status record set, the industrial state thinking is decomposed and layered according to the method described in steps S120 to S127 to generate the corresponding sample industrial thinking decomposition transmission chain sampleChain. The sampleChains corresponding to all sample industrial operation status record sets constitute the sample industrial thinking decomposition transmission chain set chainSet.

[0075] Step S214: Each sample industrial thinking decomposition and transmission chain in the sample industrial thinking decomposition and transmission chain set is sequentially input into the industrial thinking chain distillation and transfer network to be trained. The industrial thinking chain distillation and transfer network to be trained outputs the corresponding sample industrial thinking transmission deduction trajectory record, and extracts the network-generated thinking transmission intermediate state description text and the network-generated transmission confidence scalar value corresponding to each sample thinking transmission step.

[0076] For each sampleChain in the chainSet, it is input into the industrial thought chain distillation transfer network (i.e., the student network) to be trained, following the methods described in steps S131 to S135. The student network outputs the corresponding sample industrial thought transmission deduction trajectory record TRAJstudent. From TRAJstudent, the network-generated thought transmission intermediate state description text TEXTstudent,p and the network-generated transmission confidence scalar value confStudent,p are extracted.

[0077] Step S215: Compare the semantic alignment of the network-generated intermediate state description texts with the sample intermediate state description texts with the same sequential number, calculate the word-level cross-entropy loss between the network-generated intermediate state description texts and the sample intermediate state description texts, and sum the word-level cross-entropy loss corresponding to each sequential number to obtain the total semantic deviation loss.

[0078] For each step p, TEXTstudent,p and TEXTteacher,p are lexicalized to obtain the token sequences tokenStudent and tokenTeacher, with lengths Ls and Lt, respectively. The cross-entropy loss CEp = -(1 / Lt) × Σlog(P(tokenTeacher[q]|tokenStudent[1:q-1])) is calculated, where P is the probability distribution of the network predicting the next token. The CEp values ​​from all P steps are summed to obtain the total text semantic deviation loss LCE = ΣCEp.

[0079] Step S216: Compare the numerical differences between the network-generated transmission confidence scalar values ​​with the same sequential number and the sample transmission confidence scalar values ​​of the corresponding sample thought transmission steps in the sample thought transmission trajectory record. Calculate the mean squared error loss between the two, and sum the mean squared error losses corresponding to each sequential number to obtain the total confidence deviation loss. Normalize both the total text semantic deviation loss and the total confidence deviation loss. Then, perform a weighted sum of the normalized total text semantic deviation loss and the normalized total confidence deviation loss to construct the total industrial thought chain distillation joint loss. Use the total industrial thought chain distillation joint loss to perform back gradient propagation update processing on all adjustable weight parameters in the industrial thought chain distillation transfer network to be trained. The back gradient propagation update processing determines the update magnitude and direction of the weight parameters based on the partial derivative of the total industrial thought chain distillation joint loss with respect to each weight parameter.

[0080] The teacher model's output of the sample thought process transmission trajectory record also includes a scalar value of the sample transmission confidence level for each step: confTeacher, p. The mean squared error loss is calculated as MSEp = (confStudent, p - confTeacher, p). 2 Accumulate the MSEp values ​​for all P steps to obtain the confidence deviation from the total loss LMSE = ΣMSEp. Normalize LCE and LMSE to obtain normmLCE = LCE / (LCE+LMSE) and normLMSE = LMSE / (LCE+LMSE). Set weighting coefficients λCE and λMSE, typically λCE = 0.7 and λMSE = 0.3. Construct the total joint loss Ltotal = λCE × normmLCE + λMSE × normLMSE. Use Ltotal to update all adjustable weight parameters θ in the student network using backpropagation. η is the learning rate, initially set to 0.0001, and the Adam optimizer is used for parameter updates.

[0081] Step S217: When the total joint loss of the industrial thinking chain distillation converges to the preset convergence condition, the corresponding weight parameters are configured as the network parameter configuration information of the pre-constructed industrial thinking chain distillation migration network.

[0082] Repeat steps S213 to S216 to perform multiple rounds of iterative training on all samples in the sample industrial thinking decomposition transfer network set. When the decrease in Ltotal is less than 0.001 for five consecutive training rounds, the total loss is determined to have converged to the preset convergence condition. At this point, the weight parameter configuration of the student network is saved as the final network parameter configuration information of the pre-constructed industrial thinking chain distillation transfer network.

[0083] Step S310: Construct an industrial thought transmission visualization map to present the thought transmission process.

[0084] This embodiment also provides a method for constructing an industrial thinking transmission visualization map, which is used to present the abstract thinking transmission process in a graphical way.

[0085] Step S311: Extract the thought transmission step number, thought transmission intermediate state description text, and transmission confidence scalar value of each thought transmission step from the industrial thought transmission deduction trajectory record in sequence to generate a transmission trajectory analysis dataset.

[0086] For each item in TRAJrecord, extract the triplet (stepIdx, TEXTdesc, confDesc) to form the parseSet of the conduction trajectory parsing dataset.

[0087] Step S312: Extract the names of industrial equipment entities and industrial process node entities contained in the description text of each intermediate state of thought transmission in the transmission trajectory parsing dataset, and use them as the set of associated entity tags for that thought transmission step.

[0088] Entity recognition is performed on each TEXTdesc in parseSet using a conditional random field model based on a bidirectional long short-term memory network to identify industrial equipment entity names (such as "R001", "C002") and industrial process node entity names (such as "rectification section" and "stripping section"). After deduplicating all identified entity names, the resulting set is used as the associated entity label set entitySet(stepIdx) for that step.

[0089] Step S313: Based on the connection relationship of the directed thinking transmission connection edges contained in the industrial thinking decomposition transmission chain, determine the transmission predecessor and successor correspondence between different thinking transmission steps, and generate a basic transmission directed graph structure with the thinking transmission step number as the node identifier and the transmission predecessor and successor correspondence as the directed edges.

[0090] The industrial thinking deconstruction transmission chain CHAINTHINK contains a directed edge EDGEl1→l2, where l1 and l2 are hierarchical position markers. Since each hierarchical position marker corresponds to a thought transmission step, the hierarchical position markers can be mapped to step numbers. A node set V = {stepIdx | stepIdx = 1, 2, ..., L} is established. For each EDGEl1→l2, a directed edge is created from node stepIdx = l1 to node stepIdx = l2. All nodes and directed edges constitute the basic transmission directed graph Gbase.

[0091] Step S314: Fill the node attribute field of the corresponding node identifier in the basic directed graph structure with the intermediate state description text, the scalar value of the transmission confidence and the set of associated entity labels corresponding to each thought transmission step, and generate an industrial thought transmission directed graph with node attribute information.

[0092] For each node `stepIdx` in the basic directed graph Gbase, its corresponding `TEXTdesc`, `confDesc`, and `entitySet(stepIdx)` are populated into the node's attribute fields. The data structure of the node attribute fields is `nodeAttr(stepIdx) = (text=TEXTdesc, confidence=confDesc, entities=entitySet(stepIdx))`. The populated directed graph is denoted as `Gattr`.

[0093] Step S315: Based on the scalar values ​​of the transmission confidence of the predecessor node and the transmission confidence of the successor node connected to each directed edge in the directed graph of industrial thinking transmission, calculate the transmission strength coefficient of the directed edge. The transmission strength coefficient is proportional to the product of the scalar values ​​of the transmission confidence of the successor node and the transmission confidence of the predecessor node.

[0094] For each directed edge in Gattr, which connects the predecessor node u and the successor node v, the conduction confidence scalar value of the predecessor node u is confu, and the conduction confidence scalar value of the successor node v is confv. The conduction strength coefficient of this directed edge is edgeStrength(edge) = confu × confv.

[0095] Step S316: For each node in the industrial thinking transmission directed graph, find the corresponding node fill color value in the preset color gradient mapping table according to the transmission confidence scalar value of the node, and determine the display size scaling ratio of the node according to the number of entities in the associated entity label set of the node.

[0096] A color gradient mapping table, `colorMap`, is pre-defined to linearly map confidence values ​​in the [0, 1] interval to the HSV color space: low confidence (close to 0) is mapped to a red hue, and high confidence (close to 1) is mapped to a green hue. For node `v`, its confidence value `confv` is found in `colorMap` to obtain the fill color value `colorv`. Simultaneously, the number of entities in the associated entity label set of node `v` is denoted as `cntEntitiesv`. A base node radius `Rbase` is set, and the display size scaling ratio `scalev = 1 + log(1 + cntEntitiesv) / log(1 + maxCnt)`, where `maxCnt` is the maximum number of entities among all nodes. The final display radius of a node is `Rv = Rbase × scalev`.

[0097] Step S317: For each directed edge in the directed graph of industrial thinking transmission, find the corresponding directed edge line width value in the preset line width gradient mapping table according to the transmission strength coefficient of the directed edge, and find the corresponding directed edge transparency value in the preset transparency gradient mapping table according to the transmission strength coefficient of the directed edge.

[0098] A predefined linewidth gradient mapping table, `widthMap`, linearly maps the conduction strength coefficient `edgeStrength` ∈ [0, 1] to the linewidth interval [Wmin, Wmax], where Wmin is 1 pixel and Wmax is 8 pixels. A predefined transparency gradient mapping table, `alphaMap`, linearly maps `edgeStrength` to the transparency interval [αmin, αmax], where αmin is 0.2 (relatively transparent) and αmax is 1.0 (opaque). For a directed edge `edge`, the linewidth value `widthEdge` and the transparency value `alphaEdge` are obtained by looking up `edgeStrength(edge)` in the table.

[0099] Step S318: Call the preset industrial vector graphics drawing interface, draw the node circular primitives according to the node identifier coordinates, node fill color values ​​and display size scaling ratio of each node in the industrial thinking transmission directed graph, draw the directed edge arrow primitives according to the connection path, directed edge drawing line width value and directed edge drawing transparency value of each directed edge in the industrial thinking transmission directed graph, and draw the node circular primitives after the node circular primitives are drawn, add the simplified summary text label of the thinking transmission intermediate state description text corresponding to the node at the adjacent position of each node circular primitive. The simplified summary text label is obtained by truncating the first preset length of characters of the thinking transmission intermediate state description text.

[0100] Calling a pre-defined industrial vector graphics drawing interface (such as a drawing engine based on the scalable vector graphics standard), for each node v in Gattr, draws a circular primitive based on its layout coordinates (xv, yv) (pre-calculated using a force-directed layout algorithm), fill color colorv, and display radius Rv. For each directed edge, calculate the connection path (Bezier curve) based on the coordinates of its predecessor node (xu, yu) and successor node (xv, yv), and draw a curved primitive with an arrowhead using width width and transparency alpha edge. After all the circular primitives for nodes are drawn, for each node v, extract the first Lchar characters of its TEXTdesc text (e.g., Lchar=20) as a simplified summary text label, and append this text label to the upper right adjacent position of the circular primitive of that node.

[0101] Step S319: The vector graphic file obtained after completing all drawing operations is output as the industrial thinking transmission visualization map. The industrial thinking transmission visualization map is used to render and display on the industrial think tank interactive terminal to present the complete logic of the industrial thinking transmission process.

[0102] The results of all drawing operations are merged into a single, scalable vector graphics file, which is then output as a visualization of industrial thinking transmission. This visualization can be sent to an industrial think tank interactive terminal for rendering and display, allowing operators to intuitively observe the complete logical framework of the thinking transmission process.

[0103] Step S410: The industrial think tank-assisted control instruction sequence is subjected to control range expansion processing to generate a group control instruction set covering multiple related industrial execution terminals.

[0104] This embodiment also provides a method for extending control commands to a group, which extends control commands for a single device to a topologically associated group of devices.

[0105] Step S411: Obtain the industrial field equipment topology connection relationship description file associated with the initial industrial operation status record set uploaded by the industrial field time sequence acquisition link. The industrial field equipment topology connection relationship description file records the physical connection path and signal transmission direction between the source devices of each industrial equipment operation parameter snapshot unit.

[0106] The industrial field device topology connection description file is a structured data file stored in a graph description language format. This file records the unique identifier of each industrial device, its device type code, its geographical coordinates, and the physical connections between devices. Each connection record includes the source device identifier, the target device identifier, and the connection type (e.g., material piping, signal cable).

[0107] Step S412: Parse the industrial field equipment topology connection relationship description file and extract the directed graph structure of the industrial field equipment topology with each industrial equipment as a node and the physical connection path between the equipment as an edge.

[0108] The topology connection description file is parsed to extract all device identifiers as a node set (nodeSet). For each connection record, a directed edge is created from the source device node to the target device node. All nodes and directed edges constitute the directed graph Gtopo of the industrial field device topology.

[0109] Step S413: Read the control instructions one by one from the industrial think tank-assisted control instruction sequence, and extract the control receptor unit identifier and control tendency descriptor for each control instruction.

[0110] The industrial think tank-assisted regulation instruction sequence cmdSeq contains multiple regulation instruction units cmdUnit. Each cmdUnit is parsed to extract its regulatory receptor unit identifier targetId and regulatory tendency descriptor actionDesc.

[0111] Step S414: Locate the source device node that matches the control receptor unit identifier in the directed graph structure of the industrial field equipment topology, and perform a topological neighborhood expansion search along the signal transmission direction starting from the source device node. Add the device identifiers corresponding to all device nodes traversed during the topological neighborhood expansion search to the set of control receptor expansion device identifiers associated with the current control command. The search depth is determined based on a preset topological transmission depth threshold.

[0112] In Gtopo, locate the node that matches `targetId` as the starting node (`startNode`). Set a topology propagation depth threshold `Dmax` (e.g., `Dmax=3`). Starting from `startNode`, perform a breadth-first search, expanding outwards along the direction of the directed edges (i.e., the signal propagation direction). For all nodes `currentNode` traversed within the search depth `d` (1≤d≤Dmax), add their corresponding device identifiers to the extended device identifier set `extSet(targetId)`. After the search is complete, `extSet(targetId)` contains the starting node itself and all its downstream associated nodes whose depth does not exceed `Dmax`.

[0113] Step S415: Based on the regulation tendency descriptor and the device type code corresponding to each device identifier in the set of regulation receptor extended device identifiers, retrieve the regulation parameter adjustment direction and regulation parameter adjustment step size applicable to the device type code from the preset regulation parameter mapping table.

[0114] The preset control parameter mapping table `paramMap` is a multidimensional table with the device type code `devType` and the control tendency descriptor `actionDesc` as the joint primary key. It retrieves two parameters: the control parameter adjustment direction `paramDir` (with a value of +1 or -1, indicating increase or decrease) and the control parameter adjustment step size `paramStep` (with a proportional coefficient, e.g., 0.05 representing a 5% adjustment). For each device identifier `extId` in `extSet(targetId)`, its device type code `type(extId)` is first queried, and then the corresponding `paramDir` and `paramStep` are retrieved from `paramMap` using `(type(extId)`, `actionDesc)` as the key.

[0115] Step S416: Combine the regulation tendency descriptor, each device identifier in the set of extended device identifiers for regulation receptors, the regulation parameter adjustment direction corresponding to the device identifier, and the regulation parameter adjustment step size into an extended regulation instruction entry.

[0116] For each extId in extSet(targetId), generate an extended control instruction entry extCmdEntry=(extId, actionDesc, paramDir, paramStep). Iterate through all device identifiers in extSet(targetId) and generate an extended control instruction entry set extCmdSet corresponding to the current original control instruction.

[0117] Step S417: Traverse all control instructions in the industrial think tank-assisted control instruction sequence, and repeatedly execute the generation of the control receptor extended device identifier set and the processing of extended control instruction entries to generate an extended control instruction entry set containing all extended control instruction entries.

[0118] For each original control instruction cmdUnit in cmdSeq, repeat steps S413 to S416 to generate its corresponding extCmdSet. Merge all the extCmdSets corresponding to the original control instructions to form a large extended control instruction entry set totalExtCmdSet.

[0119] Step S418: Perform deduplication and merging on the extended regulatory instruction entries with the same regulatory receptor unit identifier in the extended regulatory instruction entry set. When multiple extended regulatory instruction entries point to the same regulatory receptor unit identifier, compare the regulatory tendency pointing descriptors of each extended regulatory instruction entry and retain the extended regulatory instruction entry with the highest priority according to the preset regulatory tendency priority rule.

[0120] Pre-defined regulatory preference priority rules: for example, TURNOFF has higher priority than DECREASE, DECREASE has higher priority than INCREASE, and INCREASE has higher priority than TURNON. Entries in totalExtCmdSet are grouped according to the regulatory receptor unit identifier extId. For multiple entries within each extId group, their actionDesc are compared, and the entry with the highest priority is retained according to the priority rules, while other entries in the group are deleted. After deduplication and merging, the final extended regulatory instruction entry set finalExtCmdSet is obtained.

[0121] Step S419: The final extended control instruction set obtained after deduplication and merging is used as the group control instruction set, and the group control instruction set is distributed to the corresponding industrial field control execution terminal to trigger the coordinated parameter adjustment of the associated equipment group.

[0122] The finalExtCmdSet is used as a set of group control instructions. For each extCmdEntry, its extId is extracted, and the network address of the corresponding target control execution terminal is found in the industrial field device registry. The control instruction content is then encapsulated and distributed according to steps S153 and S154 to trigger the coordinated parameter adjustment of the associated device group.

[0123] Step S510: Perform time-series segmentation processing on the initial industrial operation status record set to form segmented industrial operation status record subsets oriented towards different time analysis granularities.

[0124] This embodiment also provides a time-series segmentation method for dividing long-term operation status records into multiple time segments for independent analysis.

[0125] Step S511: Read the collection timestamp mark value carried by each industrial equipment operation parameter snapshot unit and each production process status snapshot unit from the initial industrial operation status record set, construct an ordered sequence of collection timestamp marks containing all collection timestamp mark values, and determine the complete time span length value of the initial industrial operation status record set based on the collection timestamp mark value at the beginning and end of the ordered sequence of collection timestamp marks.

[0126] Extract each from Dinit and Carrying collection timestamp marker value ,according to Construct an ordered sequence of collected timestamps Set the first timestamp. End timestamp Length of the complete time span .

[0127] Step S512: Calculate the total number of segmented truncation windows based on the preset time segmentation granularity indicator parameter and the complete time span length value. The preset time segmentation granularity indicator parameter is used to specify the time length specification of each segmented truncation window.

[0128] The preset time segmentation granularity indicator parameter G represents the time length of each segmented window (e.g., G = 3600 seconds). The total number of segmented windows Nwin = ceil(Ttotal / G), where ceil is the floor function.

[0129] Step S513: Starting from the beginning of the ordered sequence of collected timestamps in chronological order, define the continuous time intervals corresponding to the time length specifications specified by the preset time segmentation granularity indicator parameter as the time range boundaries of the current segmentation window.

[0130] For the nth segmented window (n=1, 2, ..., Nwin), its time range boundary is: The right edge of the last window may exceed... At this point, cut it to .

[0131] Step S514: Collect all industrial equipment operation parameter snapshot units and production process status snapshot units whose collected timestamp values ​​fall within the time range boundary of the current segmented interception window into a current segmented industrial operation status record subset, and assign a unique segmented window sequence number to the current segmented industrial operation status record subset.

[0132] For the nth segment window, filter out all that satisfy... of These are grouped into a subset Dseg,n of segmented industrial operation status records. A sequential number n is assigned to this subset.

[0133] Step S515: Determine whether the time range boundary of the current segmented truncation window has covered the end of the ordered sequence of the collected timestamps. If it has not covered, move the segmented truncation window to the next continuous time interval and repeat the aggregation process until the segmented truncation window covers all the collected timestamp marker values.

[0134] Check the right boundary of the current window Is it greater than or equal to? If yes, then the segmentation is complete; if not, then increment n by 1 and return to step S514 to continue processing the next window.

[0135] Step S516: All current segmented industrial operation status record subsets arranged sequentially according to the segmented window numbering are combined to form a time-series segmented industrial operation status record set. For each segmented industrial operation status record subset in the time-series segmented industrial operation status record set, the operation of decomposing and layering the initial industrial operation status record set into industrial state thinking is executed independently, generating a segmented industrial thinking decomposition transmission chain uniquely corresponding to that segmented industrial operation status record subset. For each segmented industrial thinking decomposition transmission chain, the operation of inputting the industrial thinking decomposition transmission chain into a pre-constructed industrial thinking chain distillation migration network is executed independently, generating a segmented industrial thinking transmission deduction trajectory record uniquely corresponding to that segmented industrial thinking decomposition transmission chain.

[0136] Arrange all Dseg,n in ascending order of n to form a time-series segmented industrial operation status record set DsegAll. For each Dseg,n, independently execute steps S120 to S127 to generate the corresponding segmented industrial thinking decomposition and transmission chain chainSeg,n. For each chainSeg,n, independently execute steps S130 to S135 to generate the corresponding segmented industrial thinking transmission and deduction trajectory record TRAJseg,n.

[0137] Step S517: All the segmented industrial thinking transmission and deduction trajectory records are spliced ​​together in the time domain according to their corresponding segmented window sequence number to form a panoramic industrial thinking transmission and deduction trajectory record stream with temporal continuity. The panoramic industrial thinking transmission and deduction trajectory record stream is used to reflect the complete picture of the evolution of the state transmission process of the industrial system over time.

[0138] All TRAJseg,n records are concatenated end-to-end according to their index n, with a time segmentation marker added between adjacent records. The concatenated complete sequence constitutes the panoramic industrial thinking transmission and deduction trajectory recording stream TRAJpanorama. This panoramic industrial thinking transmission and deduction trajectory recording stream can be used to analyze the long-term trends and abrupt changes in the state transmission logic of industrial systems over time.

[0139] Step S610: Perform simulation evaluation of the effectiveness of the industrial think tank-assisted control instruction sequence.

[0140] This embodiment also provides a simulation evaluation method for the effectiveness of control commands, which is used to predict the impact of control commands on the industrial site before they are actually issued.

[0141] Step S611: Obtain the set of historical control instructions for the industrial site that is synchronously collected with the initial industrial operation status record set uploaded by the industrial site time-series acquisition link. The set of historical control instructions for the industrial site includes historical control instruction content entries and historical industrial operation status change record entries collected after the historical control instructions are executed.

[0142] For example, extract a set of historical control command records (histCmdSet) from the industrial site's historical database, covering the past year. Each entry in this set contains the content of a historical control command (cmdHist) and a record of the industrial operating status change collected after the command was executed (changeHist).

[0143] Step S612: Extract the correspondence between each pair of historical control command content entries and historical industrial operation status change record entries from the set of historical control command records of the industrial site, construct a set of historical control status change mapping pairs, perform command feature vectorization processing on each historical control command content entry in the set of historical control status change mapping pairs, extract the control tendency descriptor feature component and the control receptor unit identifier feature component contained in the historical control command content entry, and concatenate the two into a historical control command feature vector.

[0144] For each pair (cmdHist, changeHist) in histCmdSet, a mapping pair is established. cmdHist is parsed to extract the regulatory tendency descriptor actionHist and the regulatory receptor unit identifier targetHist. actionHist is one-hot encoded (dimension equal to the total number of action types Anum), and targetHist is one-hot encoded (dimension equal to the total number of devices Tnum). The two one-hot vectors are then concatenated to obtain the historical regulatory instruction feature vector fCmdHist, with a dimension of Anum + Tnum.

[0145] Step S613: Perform state change feature vectorization processing on the historical industrial operation state change record entries corresponding to each historical control instruction content entry in the historical control state change mapping set, extract the parameter offset direction change component and parameter offset magnitude change component of each monitoring parameter in the historical industrial operation state change record entry, perform standardization processing on the parameter offset magnitude change component of each monitoring parameter, and concatenate the standardized parameter offset magnitude change component and the parameter offset direction change component of each monitoring parameter to form the historical state change feature vector.

[0146] For `changeHist`, extract the changes in each monitored parameter before and after the execution of the instruction. Let the parameter value before execution be `eBefore`, and after execution be `eAfter`. The parameter offset direction change component `dirChange` = `sign(eAfter - eBefore)`. The parameter offset magnitude change component `magChange` = `|eAfter - eBefore| / range`. Calculate `(dirChange, magChange)` for all A monitored parameters, combine `dirChange` into a vector `vDirChange` (dimension A), and combine `magChange` into a vector `vMagChange` (dimension A). Then concatenate `vDirChange` and `vMagChange` to obtain the historical state change feature vector `fChangeHist`, with dimension 2A.

[0147] Step S614: Using all the historical control command feature vectors and all the historical state change feature vectors to form a training data pair set, a preset industrial state change prediction network is called to perform supervised learning training on the set based on the training data, and the industrial state change prediction network after training is obtained.

[0148] The industrial state change prediction network adopts a three-layer fully connected network structure. The input layer has a dimension of Anum + Tnum, the hidden layer dimensions are 256 and 128 respectively, and the output layer dimension is 2A. A linear rectified function is used as the activation function, while no activation function is used in the output layer. (fCmdHist, fChangeHist) are used as training data pairs, and mean squared error is used as the loss function. The Adam optimizer is used for training with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. After training, the industrial state change prediction network modelPredict is obtained.

[0149] Step S615: Perform feature vectorization on each control instruction in the industrial think tank-assisted control instruction sequence to obtain the control instruction feature vector to be simulated corresponding to the control instruction, and input it into the industrial state change prediction network after training to calculate the simulation state change feature vector corresponding to the control instruction feature vector to be simulated.

[0150] For each control instruction cmdUnit in cmdSeq, it is converted into a control instruction feature vector fCmdSim to be simulated according to the method in step S612. fCmdSim is input into modelPredict, and the network forward propagation calculates the output to obtain the simulation state change feature vector fChangeSim.

[0151] Step S616: Analyze the simulation state change feature vector, decompose the simulation parameter offset direction change component and simulation parameter offset magnitude change component for each monitoring parameter, and denormalize the simulation parameter offset magnitude change component of each monitoring parameter according to the standardized parameters used when training the industrial state change prediction network, restore the expected change amount with the same physical unit and dimension as the monitoring parameters in the initial industrial operation state record set, and then superimpose the expected change amount with the corresponding current parameter offset state in the initial industrial operation state record set to generate a simulation expected state distribution description of each monitoring parameter after executing the control command.

[0152] Extract the simulation parameter offset direction change component vDirSim from the first A components of fChangeSim, and extract the simulation parameter offset magnitude change component vMagSim (this value is on the standardized scale) from the last A components. Record the mean μmag and standard deviation σmag used to standardize magChange during training. For each monitored parameter a, destandardize to obtain the expected change Δea, sim = vMagSim[a] × σmag + μmag. Let the value of this parameter at the current time be eCurrent, then the simulation expected value after executing the instruction is eSim = eCurrent + Δea, sim. Combine the simulation expected values ​​of all monitored parameters to obtain the simulation expected state distribution description stateSim.

[0153] Step S617: Summarize the simulation expected state distribution descriptions corresponding to all control commands in the industrial think tank-assisted control command sequence, and generate control command effectiveness simulation evaluation results. The control command effectiveness simulation evaluation results are used to determine the impact trend of control actions on the industrial site state in advance before the actual issuance of control commands.

[0154] Repeat steps S615 and S616 for each instruction in cmdSeq to generate the corresponding stateSim. Arrange all stateSims in the order of instruction execution to form the simulation evaluation result simResult of the control instruction effectiveness. This evaluation result can be used to pre-determine the impact trend of control actions on the industrial field status before actually issuing control instructions, assisting operators in making decisions.

[0155] Step S710: Perform cross-time correlation analysis on the industrial thinking transmission and deduction trajectory records acquired in multiple consecutive time periods.

[0156] This embodiment also provides a cross-time period correlation analysis method to identify the evolution pattern of the transmission logic of the operating status of industrial systems in different time periods.

[0157] Step S711: The industrial thinking transmission and deduction trajectory records obtained from multiple consecutive time periods are sorted according to the order of collection of time periods to generate a time-series industrial thinking transmission and deduction trajectory record sequence. For each industrial thinking transmission and deduction trajectory record, the intermediate state description text of each thinking transmission step and the sequence number of the thinking transmission step are extracted. The extracted intermediate state description text is processed by text semantic vectorization representation. The intermediate state description text of the thinking transmission is mapped into a fixed-length industrial semantic representation vector using a preset industrial semantic encoding model.

[0158] The industrial thinking transmission and deduction trajectory records TRAJq (q=1, 2, ..., Q) obtained from Q consecutive time periods are arranged in chronological order to generate a time-series industrial thinking transmission and deduction trajectory record sequence TRAJseq. For each TRAJq in TRAJseq, the intermediate state description text TEXTq,p and the thought transmission step number p of each step are extracted. The preset industrial semantic encoding model encoderSem is called. This model adopts a Transformer-based encoder architecture, containing 6 encoder layers. Each encoder layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer. TEXTq,p is input into encoderSem, and the output vector corresponding to the [CLS] marker bit of the last layer of the encoder is taken as the fixed-length industrial semantic representation vector vSem,q,p of the text, with a vector dimension of 768.

[0159] Step S712: The industrial semantic representation vectors corresponding to the sequence numbers of each thought transmission step in the same industrial thought transmission deduction trajectory record are spliced ​​together in ascending order of thought transmission step numbers to generate a single transmission semantic representation sequence vector corresponding to the industrial thought transmission deduction trajectory record. All single transmission semantic representation sequence vectors are arranged in the order of collection by time period to generate a time-series single transmission semantic representation sequence vector set. The vector difference comparison is performed on the single transmission semantic representation sequence vectors of any two adjacent time periods, and the vector cosine distance between the two is calculated as the transmission semantic offset metric value between adjacent time periods.

[0160] For each TRAJq, the industrial semantic representation vectors vSem,q,1, vSem,q,2,..., vSem,q,Pq of all its steps are concatenated in ascending order of step number p to form a single-transmission semantic representation sequence vector vSeq,q with dimension 768×Pq. All vSeq,q are arranged in order of q to obtain the temporal single-transmission semantic representation sequence vector set vSeqSet. For each pair of adjacent time intervals (q, q+1), the cosine distance between vSeq,q and vSeq,q+1 is calculated. The cosine distance value is used as the offset metric for the semantic shift between adjacent time periods, with the offset metric ranging from -1 to 1. The smaller the value, the greater the difference in semantic shift between the two time periods.

[0161] Step S713: Record the adjacent time period transmitted semantic offset metric values ​​corresponding to all adjacent time period combinations, and connect all adjacent time period transmitted semantic offset metric values ​​in time period order to form a transmitted semantic offset time series curve. Identify the time node position where the transmitted semantic offset metric values ​​of adjacent time periods change abruptly in the transmitted semantic offset time series curve. Wherein, abrupt change means that the difference between the transmitted semantic offset metric values ​​of adjacent time periods exceeds the preset offset change judgment threshold.

[0162] Connect all offsets q in the order q=1, 2, ..., Q-1 to form a transitive semantic offset time-series curve curveOffset. Calculate the difference between adjacent offset metrics Δoffsetq=|offsetq+1-offsetq|. Set a threshold for offset mutation threshMut. Iterate through q, and when Δoffsetq>threshMut, mark the time node position q+1 as the mutation node position mutPos.

[0163] Step S714: Extract the intermediate state description text of the thinking transmission from the industrial thinking transmission deduction trajectory record corresponding to the time node position before and after, and compare and analyze the changes in the names of industrial equipment entities and industrial process node entities in the intermediate state description text of thinking transmission before and after the mutation occurs.

[0164] For each mutation node position mutPos, extract all intermediate state description texts of thought transmission in the pre-mutation period TRAJmutPos-1 and the post-mutation period TRAJmutPos. Using the entity recognition method in step S312, extract the sets of industrial equipment entity names entSetBefore and entSetAfter contained in the two sets of texts respectively. Calculate the symmetric difference between the two sets symEnt=(entSetBefore\entSetAfter)∪(entSetAfter\entSetBefore), which reflects the changes in entity names before and after the mutation.

[0165] Step S715: The transmission semantic offset time sequence curve, the time node position of the mutation, and the changes in the names of industrial equipment entities and industrial process node entities before and after the mutation are merged and recorded as cross-time period correlation analysis processing results. The cross-time period correlation analysis processing results are used to reflect the transition time points and transition causes of the transmission logic of the industrial system operation status in different time periods.

[0166] The curveOffset, all mutPos, and the symEnt corresponding to each mutPos are associated and stored to form the crossResult, a cross-time period correlation analysis result. This result can be used to diagnose the time points when the operational status transmission logic of an industrial system changes between different time periods, and to infer the possible causes of the change (such as equipment replacement, process change, etc.) by analyzing changes in entity names.

[0167] Step S810: Receive the feedback status record after the control command is executed, and perform control effect deviation analysis and template adaptive correction.

[0168] This embodiment also provides a method for deviation analysis of control effect and adaptive correction of template, which is used to optimize the preset industrial thinking decomposition template based on actual control feedback.

[0169] Step S811: Receive the industrial field control execution feedback status record set uploaded by the industrial field control execution terminal after executing the industrial think tank assisted control command sequence. The industrial field control execution feedback status record set includes the actual execution completion status mark of each control command and the feedback sequence of industrial equipment operating parameter snapshot units re-collected after the control command is executed.

[0170] After the industrial think tank-assisted control command sequence cmdSeq is issued and executed, the industrial field control execution feedback status record set fbSet is received from the control execution terminal. fbSet contains the actual execution completion status flag flagSuccess (a boolean value, 1 indicates successful execution and 0 indicates execution failure) for each command, as well as the feedback sequence eFb(t) of the industrial equipment operating parameter snapshot unit re-collected after the command execution.

[0171] Step S812: Select control commands marked as successfully executed from the set of industrial field control execution feedback status records, and extract the industrial equipment operating parameter snapshot unit feedback sequence re-collected after the execution of the control command corresponding to the successfully executed control command as the effective feedback status sequence.

[0172] Select control instructions with flagSuccess=1 from fbSet to form a success instruction set succCmdSet. For each instruction in succCmdSet, extract its corresponding eFb(t) as the valid feedback state sequence fbValid.

[0173] Step S813: comparing the effective feedback state sequence with the industrial regulation decision element combination for the target industrial regulation object determined according to the industrial thinking conduction deduction trajectory record, calculating the consistency ratio between the parameter offset direction of each monitoring parameter in the effective feedback state sequence and the regulation tendency direction descriptor in the industrial regulation decision element combination, determining whether the consistency ratio is lower than a preset regulation effect conformity determination threshold, and if the consistency ratio is lower than the preset regulation effect conformity determination threshold, marking the regulation instruction as a regulation effect deviation instruction.

[0174] For each instruction in succCmdSet, the corresponding industrial regulation decision element combination comprises a regulation tendency direction descriptor actionDesc. An actual change direction of each monitoring parameter before and after the instruction is executed is extracted from fbValid. Let the parameter value before the instruction is executed be eBefore, and the parameter value after the instruction is executed be eAfter, then the actual change direction dirActual=sign(eAfter-eBefore). An expected change direction dirExpected is obtained through mapping from actionDesc (for example, actionDesc=INCREASE corresponds to +1, and DECREASE corresponds to -1). A proportion of dirActual consistent with dirExpected among all the monitoring parameters is counted as a consistency ratio consRate. A regulation effect conformity determination threshold threshCons is set (for example, threshCons=0.7). If consRate<threshCons, the regulation instruction is marked as a regulation effect deviation instruction devCmd.

[0175] Step S814: collecting all regulation instruction contents corresponding to the instructions marked as regulation effect deviation instructions, thinking conduction step context information corresponding to the regulation instruction in the industrial thinking conduction deduction trajectory record, and the effective feedback state sequence to generate a regulation effect deviation analysis sample set; for each sample in the set, extracting conduction connecting edges with conduction intensity values lower than a preset threshold in conduction intensity distribution records in the thinking conduction step context information of the sample, and taking two industrial thinking atomic state nodes connected by the conduction connecting edge as candidate conduction node pairs.

[0176] Collect the instruction content cmdDev corresponding to all devCmd, the thought transmission step context information ctxDev corresponding to the instruction in TRAJrecord (including the transmission intensity distribution records of the two steps before and after the step), and fbValid to form the control effect deviation analysis sample set devSampleSet. Set a low threshold for transmission intensity, threshLow. For each sample in devSampleSet, traverse the transmission intensity distribution records in its ctxDev and find all transmission connection edges edgeLow with transmission intensity values ​​lower than threshLow. Treat the predecessor industrial thought atomic state node nodePred and the successor industrial thought atomic state node nodeSucc connected to each edgeLow as a candidate transmission node pair candPair.

[0177] Step S815: Count the frequency of occurrence of each candidate conduction node pair in the sample set of the deviation analysis of the regulation effect, and calculate the proportion of occurrence of each candidate conduction node pair in the sample set of the deviation analysis of the regulation effect; sort all candidate conduction node pairs from high to low according to the proportion of occurrence to generate a candidate conduction node pair sequence list.

[0178] Iterate through all samples in devSampleSet and count the total frequency freqPair for each candPair. The total number of samples in devSampleSet is Ndev. Calculate the occurrence ratio propPair = freqPair / Ndev for each candPair. Sort all candPairs in descending order of propPair to generate a candidate transmission node pair sequence list rankPair.

[0179] Step S816: Select a preset number of candidate transmission node pairs with the highest frequency of occurrence from the candidate transmission node pair sequence list; calculate the occurrence ratio of each selected candidate transmission node pair in the control effect deviation analysis sample set; adjust the initial thought transmission intensity coefficient of the directed thought transmission connection edge corresponding to the selected candidate transmission node pair in the preset industrial thought decomposition template by lowering the adjustment range, which is positively correlated with the occurrence ratio of the selected candidate transmission node pair; update the preset industrial thought decomposition template using the adjusted initial thought transmission intensity coefficient, so that when the industrial thought decomposition transmission chain is subsequently generated based on the updated preset industrial thought decomposition template, the transmission intensity configuration of the corresponding directed thought transmission connection edge adapts to the control response characteristics of the actual industrial site.

[0180] Select the top Msel (e.g., Msel=10) candidate transmission node pairs from rankPair, denoted as selectedPairs. For each selPair in selectedPairs, its occurrence ratio is propSel. In the preset industrial thinking decomposition template TARGET TEMPLATE, locate the directed thinking transmission connection edge edgeCorr corresponding to selPair and obtain its current initial thinking transmission strength coefficient coefOld. Adjust the correction coefficient decFactor=1-γ×propSel, where γ is a preset adjustment strength factor (e.g., γ=0.5). The updated transmission strength coefficient coefNew=coefOld×decFactor. Write coefNew back to the attribute of the corresponding edge in TARGET TEMPLATE. Repeat the above operation to update the transmission strength coefficients corresponding to all selectedPairs. The updated TARGET TEMPLATE is saved and used for subsequent rounds of industrial state thinking decomposition layered processing, so that the configuration of causal transmission paths can adaptively fit the control response characteristics of the actual industrial site.

[0181] In one exemplary embodiment, a data analysis system for industrial think tank decision support is provided. This data analysis system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2 As shown, this data analysis system for industrial think tank decision support includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a data analysis method for industrial think tank decision support. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the casing of a data analysis system used for industrial think tank decision support, or an external keyboard, touchpad, or mouse, etc.

[0182] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A data analysis method applied to industrial think tank decision support, characterized in that, The method includes: Acquire the initial industrial operation status record set uploaded by the industrial site time-series acquisition link. The initial industrial operation status record set includes industrial equipment operation parameter snapshot units and production process status snapshot units arranged in the order of acquisition time. The initial industrial operation state record set is subjected to industrial state thinking decomposition and hierarchical processing. The state offset information in the industrial equipment operation parameter snapshot unit and the production process state snapshot unit is matched with the preset industrial thinking decomposition template for causal transmission path matching to generate an industrial thinking decomposition transmission chain with a hierarchical causal transmission path. The industrial thinking decomposition transmission chain contains multiple industrial thinking atomic state nodes connected in the order of thinking deduction. The industrial thinking decomposition transmission chain is input into a pre-constructed industrial thinking chain distillation migration network. The transmission strength between nodes is calculated node by node along the thinking deduction order of the industrial thinking atomic state nodes, and the intermediate state description of the transmission deduction is generated to obtain the industrial thinking transmission deduction trajectory record. The industrial thinking transmission deduction trajectory record includes the intermediate state description of each step of thinking transmission and the distribution record of thinking transmission strength. Based on the description of intermediate states of thought transmission and the distribution of thought transmission intensity in the industrial thought transmission trajectory record, the combination of industrial control decision elements for the target industrial regulation object is determined. The combination of industrial control decision elements includes a control tendency descriptor and a control receptor unit identifier. An industrial think tank-assisted control instruction sequence is generated based on the control tendency descriptor and the control receptor unit identifier. The industrial think tank-assisted control instruction sequence is used to drive the industrial field control execution terminal to change the configuration status of operating parameters.

2. The data analysis method for industrial think tank-assisted decision-making according to claim 1, characterized in that, The step of performing industrial state thinking decomposition and hierarchical processing on the initial industrial operation state record set, matching the state offset information in the industrial equipment operation parameter snapshot unit and the production process state snapshot unit with the preset industrial thinking decomposition template to generate an industrial thinking decomposition transmission chain with a hierarchical causal transmission path, includes: Extract the parameter offset direction indicator of each monitored parameter in each industrial equipment operation parameter snapshot unit in the initial industrial operation status record set, and combine all parameter offset direction indicators of the same industrial equipment operation parameter snapshot unit into an equipment parameter offset tendency direction vector; and extract the process stagnation degree indicator of each process node in each production process status snapshot unit in the initial industrial operation status record set, and combine all process stagnation degree indicators of the same production process status snapshot unit into a process stagnation tendency direction vector. Calculate the first direction matching degree sequence between the equipment parameter offset tendency direction vector and each preset industrial thinking disassembly template, and simultaneously calculate the second direction matching degree sequence between the process blockage tendency direction vector and each preset industrial thinking disassembly template. Normalize the matching degree values ​​in the first and second direction matching degree sequences, and perform weighted fusion processing on each matching degree value in the normalized first direction matching degree sequence and the matching degree value in the normalized second direction matching degree sequence corresponding to the same order to generate a comprehensive matching degree ranking queue. Select the preset industrial thinking disassembly template corresponding to the first-ranked comprehensive matching degree from the comprehensive matching degree ranking queue as the target industrial thinking disassembly template. The original observed values ​​of each monitoring parameter contained in the industrial equipment operation parameter snapshot unit are compared one by one with the preset initial offset threshold range in the target industrial thinking decomposition template. Based on the comparison results, the hierarchical position mark of each monitoring parameter in the thinking decomposition starting level is determined. The original state description values ​​of each process node contained in the production process state snapshot unit are compared one by one with the preset initial process threshold range in the target industrial thinking decomposition template. Based on the comparison results, the hierarchical position mark of each process node in the thinking decomposition starting level is determined. For all monitoring parameter offset direction indicators and all process obstruction degree indicators belonging to the same thinking decomposition level, offset polarity aggregation statistical processing is performed. The mode direction of the offset direction indicators within the thinking decomposition level is extracted as the offset direction aggregation representation. The offset degree value of each type of monitoring parameter within the thinking decomposition level is normalized. Then, the weighted average of the normalized offset degree values ​​of all monitoring parameters is calculated to obtain the comprehensive offset magnitude aggregation representation of the thinking decomposition level. The weights are determined according to the preset importance coefficient of each monitoring parameter in the preset industrial thinking decomposition template. The offset direction aggregation representation and the offset magnitude aggregation representation are jointly constructed to form the industrial thinking atomic state node corresponding to the thinking decomposition level. The industrial thinking atomic state node simultaneously contains the hierarchical position mark, the offset direction aggregation representation and the offset magnitude aggregation representation. Based on the pre-established hierarchical causal transmission dependency path within the target industrial thinking decomposition template, directed thinking transmission connection edges are established between industrial thinking atomic state nodes with different hierarchical position marks. The direction of the directed thinking transmission connection edges is from the industrial thinking atomic state node with the earlier hierarchical position mark to the industrial thinking atomic state node with the later hierarchical position mark. All industrial thinking atomic state nodes and all directed thinking transmission connection edges are merged to form the industrial thinking decomposition transmission chain with hierarchical causal transmission path.

3. The data analysis method for industrial think tank-assisted decision-making according to claim 1, characterized in that, The process involves inputting the decomposed industrial thinking transmission chain into a pre-constructed industrial thinking chain distillation migration network, calculating the inter-node transmission strength along the thought deduction order of the industrial thinking atomic state nodes, generating intermediate state descriptions of transmission deduction, and obtaining a record of the industrial thinking transmission deduction trajectory, including: The industrial thinking atomic state nodes contained in the hierarchical causal transmission chain are vectorized. The offset direction aggregation representation and offset magnitude aggregation representation contained in each industrial thinking atomic state node are jointly mapped into a fixed-length industrial thinking node embedding vector, resulting in a fixed-length industrial thinking node embedding vector sequence that corresponds one-to-one with all industrial thinking atomic state nodes. A pre-trained industrial thinking start embedding vector is inserted at the beginning of the fixed-length industrial thinking node embedding vector sequence, and a pre-trained industrial thinking end embedding vector is inserted at the end of the fixed-length industrial thinking node embedding vector sequence, generating a complete industrial thinking transmission embedding vector sequence. Add a position encoding vector corresponding to its position in the sequence to each embedding vector in the complete industrial thinking transmission embedding vector sequence to obtain an industrial thinking transmission embedding vector sequence carrying position encoding. The numerical distribution of the position encoding vector is generated according to a preset sine-cosine alternation function. The industrial thinking transmission embedding vector sequence carrying the position code is sent to the first transmission deduction step of the industrial thinking chain distillation transfer network. In the first transmission deduction step, for each industrial thinking node embedding vector in the sequence, the vector dot product similarity between the industrial thinking node embedding vector and all industrial thinking node embedding vectors arranged in the sequence that are themselves and before it is calculated. The calculated vector dot product similarity set is numerically scaled through a preset normalization exponential function to obtain the attention weight distribution value corresponding to each of the industrial thinking node embedding vectors arranged in the sequence that are themselves and before it. The attention weight distribution value is used to perform a weighted summation on all industrial thinking node embedding vectors arranged in the sequence that are themselves and before it to generate the updated embedding vector after the first transmission deduction step. After traversing all industrial thinking node embedding vectors, the updated embedding vector sequence of the first transmission deduction step is obtained. The first transmission deduction step update embedding vector sequence is used as the input of the second transmission deduction step. The vector dot product similarity calculation, attention weight distribution value calculation and weighted summation are repeatedly performed to generate the second transmission deduction step update embedding vector sequence. This process is repeated until the preset total number of transmission deduction steps is completed to obtain the final transmission deduction update embedding vector sequence. The starting position update embedding vector corresponding to the industrial thinking starting embedding vector is extracted from it. The starting position update embedding vector is input into the transmission intermediate state description generator of the industrial thinking chain distillation transfer network. According to the input starting position update embedding vector, the corresponding word is retrieved one by one in the preset industrial thinking transmission description vocabulary and output. The retrieved word is concatenated in order to form the thinking transmission intermediate state description corresponding to each thinking transmission step. At the same time as outputting each thinking transmission intermediate state description, the transmission confidence scalar value corresponding to the thinking transmission intermediate state description is calculated. The intermediate state description and confidence scalar value of each step of the thought transmission are associated and stored in the order of the thought transmission steps to generate the industrial thought transmission deduction trajectory record. Each record in the industrial thought transmission deduction trajectory record includes the thought transmission step number, the intermediate state description text, and the confidence scalar value.

4. The data analysis method for industrial think tank-assisted decision-making according to claim 1, characterized in that, The step of determining the combination of industrial regulation decision-making elements for the target industrial regulation object based on the description of intermediate states of thought transmission and the distribution record of thought transmission intensity in the industrial thought transmission deduction trajectory record includes: The intermediate state description text and the scalar value of the transmission confidence of each step of the industrial thought transmission deduction trajectory are extracted sequentially from the record. The intermediate state description text and the scalar value of the transmission confidence of the same thought transmission step number are associated to form a single-step transmission analysis unit. The intermediate state description text of the thought transmission in each single-step transmission analysis unit is split into regulatory semantic components to obtain the regulatory action semantic fragment and regulatory receptor semantic fragment contained in the intermediate state description text. The regulatory action semantic fragment is matched and converted with a preset set of regulatory action semantic mapping rules to generate the regulatory tendency pointing descriptor corresponding to the single-step transmission analysis unit. The regulatory tendency pointing descriptor is used to characterize the type of industrial regulation operation direction. The semantic fragment of the regulatory receptor is matched and converted with a set of preset industrial entity identifier mapping rules to generate the regulatory receptor unit identifier corresponding to the single-step transmission parsing unit. The regulatory receptor unit identifier is consistent with the unique identifier code of the source equipment of the industrial equipment operation parameter snapshot unit or the source process node of the production process status snapshot unit. The conduction confidence scalar value of the single-step conduction analysis unit is used as a weighting factor for the corresponding regulatory tendency pointing descriptor and regulatory receptor unit identifier. The regulatory tendency pointing descriptor, regulatory receptor unit identifier and weighting factor are combined into single-step regulation candidate items. In the single-step regulation candidate item set, for items with the same regulatory receptor unit identifier, the cumulative score of each regulatory tendency pointing descriptor is calculated by weighting the weight factor. The regulatory tendency pointing descriptor with the highest cumulative score is paired with the regulatory receptor unit identifier as the target regulation decision element. All target regulation decision elements are collected to form the industrial regulation decision element combination for the target industrial regulation object.

5. The data analysis method for industrial think tank-assisted decision-making according to claim 1, characterized in that, The generation of the industrial think tank-assisted regulation instruction sequence based on the regulation tendency descriptor and the regulation receptor unit identifier includes: Read the target control decision elements one by one from the combination of industrial control decision elements, obtain the control tendency descriptor and control receptor unit identifier contained in each target control decision element, and retrieve the corresponding set of controllable parameter names and the current operating setting value of each controllable parameter name from the preset industrial control parameter configuration library based on the control receptor unit identifier. For each adjustable parameter name, the target adjustment operation to be performed on the current operating setting value of the adjustable parameter name is determined according to the adjustment direction indicated by the adjustment tendency descriptor, so as to obtain the target operating setting value of the adjustable parameter name; The regulatory receptor unit identifier, the name of each adjustable parameter and its target operating setting value are combined into a single regulatory instruction content entry. Obtain the terminal addressing identifier of the target control execution terminal that is pre-bound to the control receptor unit identifier, encapsulate the control instruction content item with the terminal addressing identifier to generate an industrial control instruction unit, and arrange all industrial control instruction units in the original order of each target control decision element in the industrial control decision element combination to form an industrial think tank-assisted control instruction sequence, and convert it into an instruction data block in a preset industrial control communication protocol format for sending to the target control execution terminal.

6. The data analysis method for industrial think tank-assisted decision-making according to claim 1, characterized in that, Before inputting the industrial thinking decomposition transmission chain into the pre-constructed industrial thinking chain distillation transfer network, an industrial thinking chain distillation pre-training process is performed to obtain the network parameter configuration information of the industrial thinking chain distillation transfer network. The industrial thinking chain distillation pre-training process includes: The sample thought transmission trajectory record output by a large language model in the industrial field for a sample industrial operation status record set is obtained. The sample thought transmission trajectory record includes a sequence of sample thought transmission steps and a sample thought transmission intermediate state description text corresponding to each sample thought transmission step. Each step of the sample thought transmission trajectory record is marked with a step number, and each sample thought transmission intermediate state description text is assigned a unique sequential number in the sample thought transmission step sequence to form a sample thought transmission supervision signal set. Using the same processing method as described above for decomposing and layering the initial industrial operation status record set, the sample industrial operation status record set is processed to generate a sample industrial thinking decomposition transmission chain set corresponding to the sample industrial operation status record set. Each sample industrial thinking decomposition and transmission chain in the sample industrial thinking decomposition and transmission chain set is sequentially input into the industrial thinking chain distillation and transfer network to be trained. The industrial thinking chain distillation and transfer network to be trained outputs the corresponding sample industrial thinking transmission deduction trajectory record, and extracts the network-generated thinking transmission intermediate state description text and the network-generated transmission confidence scalar value corresponding to each sample thinking transmission step. The network-generated intermediate state description texts with the same sequential number are compared with the sample intermediate state description texts for text semantic alignment. The word-level cross-entropy loss between the network-generated intermediate state description texts and the sample intermediate state description texts is calculated, and the word-level cross-entropy loss corresponding to each sequential number is accumulated to obtain the total text semantic deviation loss. The network-generated transmission confidence scalar value with the same sequential number is compared with the sample transmission confidence scalar value of the corresponding sample thought transmission step in the sample thought transmission trajectory record. The mean squared error loss between the two is calculated, and the mean squared error loss corresponding to each sequential number is accumulated to obtain the total confidence deviation loss. The total text semantic deviation loss and the total confidence deviation loss are normalized respectively. Then, the normalized total text semantic deviation loss and the normalized total confidence deviation loss are weighted and summed to construct the total industrial thought chain distillation joint loss. The total industrial thought chain distillation joint loss is used to perform back gradient propagation update processing on all adjustable weight parameters in the industrial thought chain distillation transfer network to be trained. The back gradient propagation update processing determines the update magnitude and direction of the weight parameters based on the partial derivative of the total industrial thought chain distillation joint loss with respect to each weight parameter. When the total joint loss of the industrial thinking chain distillation converges to the preset convergence condition, the corresponding weight parameters are configured as the network parameter configuration information of the pre-constructed industrial thinking chain distillation migration network.

7. The data analysis method for industrial think tank-assisted decision-making according to claim 1, characterized in that, The method further includes: The thought transmission step number, thought transmission intermediate state description text, and transmission confidence scalar value of each thought transmission step are extracted sequentially from the industrial thought transmission deduction trajectory record to generate a transmission trajectory analysis dataset. The industrial equipment entity name and industrial process node entity name contained in each thought transmission intermediate state description text in the transmission trajectory analysis dataset are extracted as the set of associated entity tags for that thought transmission step. Based on the connection relationship of the directed thinking transmission connection edges contained in the industrial thinking deconstruction transmission chain, the transmission predecessor and successor correspondence between different thinking transmission steps is determined, and a basic transmission directed graph structure is generated with the thinking transmission step number as the node identifier and the transmission predecessor and successor correspondence as the directed edges. The intermediate state description text, transmission confidence scalar value, and associated entity label set corresponding to each thought transmission step are filled into the node attribute field of the corresponding node identifier in the basic transmission directed graph structure to generate an industrial thought transmission directed graph with node attribute information. Based on the transmission confidence scalar value of the predecessor node and the transmission confidence scalar value of the successor node connected by each directed edge in the industrial thought transmission directed graph, the transmission strength coefficient of the directed edge is calculated. The transmission strength coefficient is proportional to the product of the transmission confidence scalar value of the successor node and the transmission confidence scalar value of the predecessor node. For each node in the industrial thinking transmission directed graph, the corresponding node fill color value is found in a preset color gradient mapping table based on the transmission confidence scalar value of the node, and the display size scaling ratio of the node is determined based on the number of entities in the associated entity label set of the node; and for each directed edge in the industrial thinking transmission directed graph, the corresponding directed edge drawing line width value is found in a preset line width gradient mapping table based on the transmission strength coefficient of the directed edge, and the corresponding directed edge drawing transparency value is found in a preset transparency gradient mapping table based on the transmission strength coefficient of the directed edge. The system calls a preset industrial vector graphics drawing interface to draw circular primitives of nodes based on the node identifier coordinates, node fill color values, and display size scaling ratio of each node in the industrial thinking transmission directed graph. It also draws arrow primitives of directed edges based on the connection paths, line widths, and transparency values ​​of each directed edge in the industrial thinking transmission directed graph. After the circular primitives of nodes are drawn, a simplified summary text tag of the intermediate state description text corresponding to that node is added to the adjacent positions of each circular primitive. The simplified summary text tag is obtained by truncating the first character of the intermediate state description text of the thinking transmission. The vector graphic file obtained after completing all drawing operations is output as the industrial thinking transmission visualization diagram. The industrial thinking transmission visualization diagram is used to render and display on the industrial think tank interactive terminal to present the complete logic of the industrial thinking transmission process.

8. The data analysis method for industrial think tank-assisted decision-making according to claim 1, characterized in that, The method further includes: Obtain an industrial field device topology connection description file associated with the initial industrial operation status record set uploaded by the industrial field time-series acquisition link. The industrial field device topology connection description file records the physical connection path and signal transmission direction between the source devices of each industrial equipment operation parameter snapshot unit. The description file of the topological connection relationship of the industrial field equipment is parsed, and the directed graph structure of the industrial field equipment topology with each industrial equipment as a node and the physical connection path between the equipment as an edge is extracted. Read the control instructions one by one from the industrial think tank-assisted control instruction sequence, and extract the control receptor unit identifier and control tendency descriptor from each control instruction; In the directed graph structure of the industrial field equipment topology, the source device node that matches the identifier of the control receptor unit is located, and a topological neighborhood expansion search is performed along the signal transmission direction starting from the source device node. The device identifiers corresponding to all device nodes traversed during the topological neighborhood expansion search are added to the set of control receptor expansion device identifiers associated with the current control command. The search depth is determined according to a preset topological transmission depth threshold. Based on the regulatory tendency descriptor and the device type code corresponding to each device identifier in the set of regulatory receptor extended device identifiers, the regulatory parameter adjustment direction and regulatory parameter adjustment step size applicable to the device type code are retrieved from the preset regulatory parameter mapping table. The regulation tendency descriptor, each device identifier in the set of extended device identifiers of the regulation receptor, the regulation parameter adjustment direction and the regulation parameter adjustment step size corresponding to the device identifier are combined into an extended regulation instruction entry. The extended control instruction entries with the same control receptor unit identifier in the extended control instruction entry set are deduplicated and merged. The final extended control instruction entry set obtained after deduplication and merging is used as the group control instruction set. The group control instruction set is then distributed to the corresponding industrial field control execution terminal to trigger the coordinated parameter adjustment of the associated equipment group.

9. The data analysis method for industrial think tank-assisted decision-making according to claim 1, characterized in that, The method further includes: Read the collection timestamp mark value carried by each industrial equipment operation parameter snapshot unit and each production process status snapshot unit from the initial industrial operation status record set, construct an ordered sequence of collection timestamp marks containing all collection timestamp mark values, and determine the complete time span length value of the initial industrial operation status record set based on the collection timestamp mark value at the beginning and end of the ordered sequence of collection timestamp marks. Based on the preset time segmentation granularity indicator parameter and the complete time span length value, the total number of segmented truncation windows is calculated. The preset time segmentation granularity indicator parameter is used to specify the time length specification of each segmented truncation window. Starting from the beginning of the ordered sequence of timestamps collected, consecutive time intervals corresponding to the time length specifications specified by the preset time segmentation granularity indicator parameter are sequentially defined as the time range boundaries of the current segmentation window. All industrial equipment operation parameter snapshot units and production process status snapshot units whose collected timestamp values ​​fall within the time range boundary of the current segmented interception window are collected into a current segmented industrial operation status record subset, and a unique segmented window sequence number is assigned to the current segmented industrial operation status record subset. Determine whether the time range boundary of the current segmented truncation window has covered the end of the ordered sequence of the collected timestamps. If it has not covered, move the segmented truncation window to the next continuous time interval and repeat the aggregation process until the segmented truncation window covers all the collected timestamp marker values. All current segmented industrial operation status records are arranged sequentially according to the segmented window numbering and constituted as a time-series segmented industrial operation status record set. For each segmented industrial operation status record subset in the time-series segmented industrial operation status record set, the operation of decomposing and layering the initial industrial operation status record set into industrial state thinking is executed independently, generating a segmented industrial thinking decomposition transmission chain that uniquely corresponds to that segmented industrial operation status record subset. For each segmented industrial thinking decomposition transmission chain, the operation of inputting the industrial thinking decomposition transmission chain into a pre-constructed industrial thinking chain distillation and migration network is executed independently, generating a segmented industrial thinking transmission and deduction trajectory record that uniquely corresponds to that segmented industrial thinking decomposition transmission chain. All the segmented industrial thinking transmission and deduction trajectory records are sequentially numbered according to their corresponding segmented windows and spliced ​​in the time domain to form a panoramic industrial thinking transmission and deduction trajectory record stream with temporal continuity. The panoramic industrial thinking transmission and deduction trajectory record stream is used to reflect the complete picture of the evolution of the state transmission process of the industrial system over time.

10. A data analysis system for industrial think tank decision support, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the data analysis method for industrial think tank decision support as described in any one of claims 1 to 8 by executing the machine-executable instructions.

Images