Garbage incinerator self-adaptive control method fused with natural language

Abstract

The invention discloses a garbage incinerator adaptive control method fused with a natural language, and the specific implementation scheme comprises the steps: S1, obtaining manual inspection information and planned maintenance information, inputting the manual inspection information, the planned maintenance information and a control instruction into a working condition description instruction generation model in combination with cues, and generating working condition description information; s2, inputting the working condition description information into a pre-action generation model to generate pre-action generation description; s3, converting the pre-action generation description into a 1024-dimensional integral fusion feature through a vectorization model; and S4, inputting the overall fusion features and field sensor data into a data-driven action model to output a control command of an execution mechanism. The method solves the problems that the model is difficult to train and the working condition adaptability is poor due to the fact that the quantization instruction feature space of the controller is very complex.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a garbage incinerator adaptive control method fusing natural language. BACKGROUND

[0002] The working condition data of the garbage incinerator is derived from sensor data, artificial field inspection data and planned operation data (overhaul, maintenance, debugging); at present, the AI-based intelligent incineration system is basically based on sensor data for modeling and control, but some working conditions cannot be sensed by sensors, and are some special working conditions generated by artificial field inspection or some planned operation. Since the description of the working condition by the inspection personnel cannot directly act on the existing control model, the operator can only completely remove the AI and perform manual operation; although the existing LLM large model has a good understanding of natural language, the generation of the quantization instruction feature space of the controller is very complex, which leads to difficulty in training the model and poor working condition adaptability. SUMMARY

[0003] Therefore, the present application provides a garbage incinerator adaptive control method fusing natural language to solve the problem that the quantization instruction feature space of the controller is very complex, which leads to difficulty in training the model and poor working condition adaptability.

[0004] The present application provides a garbage incinerator adaptive control method fusing natural language, comprising:

[0005] S1: obtaining artificial inspection information and planned maintenance information, inputting the artificial inspection information, the planned maintenance information and a control instruction combined with a prompt word into a working condition description instruction generation model to generate working condition description information;

[0006] S2: inputting the working condition description information into a pre-action generation model to generate a pre-action generation description;

[0007] S3: converting the pre-action generation description into a 1024-dimensional overall fusion feature through a vectorization model;

[0008] S4: inputting the overall fusion feature and the sensor data on site into a data-driven action model to output a control command of an execution mechanism.

[0009] The operation of inputting the artificial inspection information and the planned maintenance information combined with the prompt word into the working condition description instruction generation model comprises,

[0010] The garbage incinerator is numbered from x1 to x n , and the numbers x1 to x n are inspected and maintained in turn.a waste incinerator, and issues a control instruction, and combines the artificial inspection information, the planned maintenance information and the control instruction into a prompt word and inputs the prompt word into a working condition description instruction generation model to perform reinforcement learning training.

[0011] The working condition description instruction generation model is trained by reinforcement learning,

[0012] According to the actual business pain points, the weights of each dimension are adjusted.

[0013] The constraint rules of the waste incineration process are coded as hard constraints of the reward function, and the instructions of the working condition description instruction generation model generated by the engineers for generating working condition description information are scored, and the working condition description information is optimized by PPO.

[0014] If the index of the running sensor exceeds the hard constraint of the reward function, a penalty will be triggered.

[0015] The hard constraint includes an SNCR reaction temperature window and an upper limit of the cooling water flow.

[0016] The operation of inputting the working condition description information into a pre-action generation model to generate a pre-action generation description includes,

[0017] The working condition description information and the sensor data in the last 1 hour are combined to generate a pre-action generation description for understanding the user's intention and the control command; wherein the sensor data adopts a sampling method of far sparse and near dense.

[0018] And use a 1.5B open source LLM model to fine-tune the pre-action generation description.

[0019] The operation of converting the pre-action generation description into a 1024-dimensional word vector feature by a vectorization model includes,

[0020] The pre-action generation description is vectorized by a vectorization model to obtain an overall fusion feature.

[0021] Based on the pre-action description vector and the current sensor data, a direct connection operation is performed to obtain an overall fusion feature.

[0022] The training of the data-driven action model includes,

[0023] The operation of organizing the historical data of the incinerator includes,

[0024] After the historical data of the incinerator is processed according to steps S1-S3, an overall fusion feature and a corresponding executed action sequence [a1, a2, a3,...a n ] are obtained every 2 seconds.

[0025] The data-driven action model specifically includes,

[0026] The linear mapping is used to preprocess the input overall fusion features;

[0027] The Transformer encoding layer is formed by stacking the Transformer modules to obtain the common features Z.

[0028] The MultiHead network based on the multi-layer perception (MLP), different heads n correspond to the final output of different actuators to perform actions, wherein the common features Z are the input of the MultiHead network.

[0029] The training of the data-driven action model comprises,

[0030] The data-driven action model is trained by using the historical data of the operation of the incinerator, so that the deviation between the action instructions output by the network and the actual execution actions reaches a minimum.

[0031] The calculation of the loss value when the deviation between the action instructions output by the network and the actual execution actions reaches a minimum comprises,

[0032] ;

[0033] Wherein, represents the loss value, represents the i-th number of the action sequence, represents the action instruction predicted by the model is the actual action instruction.

[0034] Beneficial effects: The application uses the experiential information of artificial natural language input to make the control model more comprehensively understand the nature of the working condition, avoid the control failure caused by incomplete information, and reduce the situation that artificial control is completely cut off.

[0035] Fill the cognitive gap between sensor data and actual working conditions, and make the control logic more suitable for the actual needs of the field;

[0036] Under complex working conditions, artificial complete takeover of control not only consumes time and effort, but also may cause risks such as furnace temperature fluctuation and excessive pollutant emission due to insufficient experience or delayed reaction. Through natural language interaction, artificial only needs to provide key information, and the model can automatically adjust the strategy, which not only reduces the labor input, but also reduces the human operation error through standardized logic optimization, and ensures the stable operation of the equipment.

[0037] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the application, nor is it used to limit the scope of the application. Other features of the application will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0038] The accompanying drawings are used to better understand the present application and do not constitute a limitation on the present application. Among them:

[0039] Figure 1 is a flowchart provided according to the present application. DETAILED DESCRIPTION

[0040] The exemplary embodiments of the present application are described below in conjunction with the accompanying drawings, which include various details of the embodiments of the present application to help understanding, and should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Also, in order to be clear and concise, the description below omits the description of well-known functions and structures.

[0041] As Figure 1 shown, the present application provides a natural language fusion garbage incinerator adaptive control method, comprising:

[0042] S1: obtaining artificial inspection information and planned maintenance information, combining the artificial inspection information, the planned maintenance information and the control instruction with the prompt word to input the working condition description instruction generation model, and generating working condition description information. It should be noted that:

[0043] The operation of inputting the artificial inspection information and the planned maintenance information combined with the prompt word into the working condition description instruction generation model includes,

[0044] The garbage incinerator is numbered from x1 to x n , and the garbage incinerator numbered from x1 to x n is inspected, maintained and controlled in turn, and the artificial inspection information, the planned maintenance information and the control instruction are combined with the prompt word and input into the working condition description instruction generation model for reinforcement learning training.

[0045] The working condition description instruction structure is as follows: [

[0047] {

[0048] cmd_src: inspection,

[0049] related_device: [x1, x2...],

[0050] describe: [“1 No. 1 furnace burns 1 section 2 columns, 3 column grate jam”],

[0051] },

[0052] {

[0053] cmd_src: maintenance,

[0054] related_deveice:[x5,x6...],

[0055] Description: ["The No. 1 boiler cooling tower is undergoing short-term maintenance; outlet temperature will not be monitored at this time."]

[0056] },

[0057] {

[0058] cmd_src: verbal,

[0059] related_deveice:[x1,x 10 ...],

[0060] Describe: ["Reduce the secondary air intake of Boiler No. 1"]

[0061] } ... ]

[0064] The specific operations for training the working condition description instruction generation model using reinforcement learning.

[0065] Adjust the weight of each dimension according to the actual business pain points (e.g., when the problem of "lime compaction" occurs frequently recently, temporarily increase the weight of "business rationality" to 50%).

[0066] The constraint rules of the waste incineration process are encoded into hard constraints of the reward function, and the working condition description information generated by the engineer's working condition description instruction generation model is scored. The working condition description information is then optimized using PPO.

[0067] First, have engineers score the instructions generated by the model (as an initial reward), then use PPO optimization to solve the problem that "the reward function is difficult to fully quantify the rationality of the business".

[0068] If the metrics of a running sensor exceed the hard constraints of the reward function, a penalty will be triggered to prevent the model from generating instructions that do not conform to the laws of physics.

[0069] The hard constraints include the SNCR reaction temperature window and the upper limit of the cooling water flow rate.

[0070] Table 1: Reward Rules

[0071] Evaluation Difficulty Reward Rules Weights Format Compliance Full compliance with JSON structure: +2 points; Missing cmd_src or related_deveice field: -5 points; related_deveice contains a number not in the device list: -3 points; 30% Semantic Accuracy Perfect match with input natural language: +3 points; Core action error (e.g., "turn down" -> "turn up"): -10 points; Device number error (e.g., "1st furnace" -> "2nd furnace"): -8 points; 40% Business Rationality Compliance with process constraints (e.g., water flow for cooling does not exceed the limit): +3 points; Violation of key constraints (e.g., instructions that cause lime to solidify): -15 points; Existence of potential risks (e.g., not considering the atomization effect of the spray gun): -5 points; 30%

[0072] S2: Input the described working condition information into the pre-action generation model to generate a pre-action generation description. It should be noted that:

[0073] The operation of inputting the working condition description information into the pre-action generation model to generate a pre-action generation description includes,

[0074] The intention of the user is understood in combination with the working condition description information and sensor data within the last 1 hour to generate a pre-action generation description; wherein the sensor data adopts a far-far-near-close sampling method;

[0075] And use 1.5B open source LLM model to fine-tune SFT for pre-action generation description.

[0076] The intention of the user is understood in combination with the working condition description information and sensor data within the last 1 hour to generate a pre-action generation description; wherein the sensor data adopts a far-far-near-close sampling method;

[0077] The pre-action generation model input format is as follows:

[0078] {

[0079] “opera_cond”: working condition description 1

[0080] “sensor_data”: [hourly average, 50-minute average, 40-minute average... 10-minute average, 9-minute average,... 1-minute average, current value]

[0081] }

[0082] The pre-action generation model output format is as follows: [

[0084] {

[0085] cmd_type: adjustment,

[0086] related_deveice: [x1, x2...],

[0087] describe: Adjust the speed of the 2, 3 column material return device of No. 1 furnace

[0088] reason: 2, 3 section grate jam, adjust the material return speed, observe whether it helps to reduce jam

[0089] },

[0090] {

[0091] cmd_src: switch,

[0092] related_deveice: [x5, x6...],

[0093] describe: Close the automatic tracking of the outlet temperature of No. 1 furnace cooling tower, and follow the current instructions

[0094] reason: no need to track outlet temperature during cooling tower maintenance

[0095] },

[0096] {

[0097] cmd_src: warning,

[0098] related_deveice: [x1, x10...],

[0099] describe: do not recommend reducing the secondary air of No. 1 furnace

[0100] reason: the current furnace temperature is too high and the oxygen content is low

[0101] },

[0102] {

[0103] cmd_src: invalid,

[0104] related_deveice: [x1, x10...],

[0105] describe: no action,

[0106] reason: no valid pre-control command is generated, if operation is needed, use more understandable instructions

[0107] } ... ]

[0110] S3: convert the pre-action generation description into a 1024-dimensional overall fusion feature through a vectorization model. It should be noted that:

[0111] The operation of converting the pre-action generation description into a 1024-dimensional word vector feature through a vectorization model includes,

[0112] vectorizing the pre-action generation description using a vectorization model to obtain an overall fusion feature;

[0113] directly connecting the pre-action description vector and the current sensor data to obtain an overall fusion feature.

[0114] S4: input the overall fusion feature and the sensor data on site into a data-driven action model to output a control command of an execution mechanism. It should be noted that:

[0115] The training of the data-driven action model includes,

[0116] The operation of sorting the historical data of the incinerator operation includes,

[0117] After the historical data of the incinerator operation is processed according to steps S1-S3, an overall fusion feature and a corresponding current executed action sequence [a1, a2, a3,...a n ] are obtained every 2 seconds.

[0118] The data-driven action model specifically comprises,

[0119] The input overall fusion feature is preprocessed by linear mapping;

[0120] The Transformer encoding layer is composed of Transformer modules to obtain the common feature Z.

[0121] The MultiHead network based on the multi-layer perception MLP, wherein different heads n correspond to the final output actions of different actuators, and the common feature Z is the input of the MultiHead network.

[0122] The training of the data-driven action model comprises,

[0123] The historical data of the incinerator operation is used to train the data-driven action model, so that the deviation between the action instructions output by the network and the actual executed actions reaches the minimum.

[0124] The calculation of the loss value when the deviation between the action instructions output by the network and the actual executed actions reaches the minimum comprises,

[0125] ;

[0126] wherein, represents the loss value, represents the i-th value of the action sequence, represents the action instruction predicted by the model is the actual action instruction.

[0127] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any change or replacement within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for fusion natural language garbage incinerator adaptive control, characterized in that, The method comprises the following steps: S1: obtaining manual inspection information and planned maintenance information, inputting the manual inspection information, the planned maintenance information, and a control instruction combined with a prompt word into a working condition description instruction generation model to generate working condition description information; S2: inputting the working condition description information into a pre-action generation model to generate a pre-action generation description; S3: converting the pre-action generation description into a 1024-dimensional overall fusion feature through a vectorization model; S4: inputting the overall fusion feature and sensor data on site into a data-driven action model to output a control command of an execution mechanism.

2. The method according to claim 1, wherein the method is characterized in that: The operation of inputting the manual inspection information and the planned maintenance information combined with the prompt word into the working condition description instruction generation model comprises, The waste incinerator from x1 to x n The waste incinerator numbered from x1 to x n is sequentially inspected and maintained, and a control instruction is issued. The artificial inspection information, planned maintenance information and control instruction are combined with prompt words and input into a working condition description instruction generation model for reinforcement learning training.

3. The method according to claim 1 or 2, characterized in that: The specific operation of performing reinforcement learning training on the working condition description instruction generation model comprises, adjusting the weight of each dimension according to the actual business pain points; encoding the constraint rules of the waste incineration process as a hard constraint of the reward function, and scoring the working condition description information generated by the working condition description instruction generation model by the engineer, and optimizing the working condition description information by PPO; If the index of the running sensor exceeds the hard constraint of the reward function, a penalty will be triggered.

4. The method according to claim 3, wherein the method further comprises: The hard constraint comprises an SNCR reaction temperature window and an upper limit of the cooling water flow.

5. The method according to claim 1 or 4, characterized in that: The operation of inputting the working condition description information into the pre-action generation model to generate a pre-action generation description comprises, generating a pre-action generation description by understanding the user's intention and the control command in combination with the working condition description information and the sensor data within the last one hour; wherein the sensor data adopts a far-sparse-near-dense sampling method; and using a 1.5B open source LLM model to fine-tune the pre-action generation description by SFT.

6. The method according to claim 5, wherein the method further comprises: The operation of converting the pre-action generation description into a 1024-dimensional word vector feature through a vectorization model comprises, vectorizing the pre-action generation description by using a vectorization model to obtain an overall fusion feature; directly connecting the pre-action description vector and the current sensor data to obtain the overall fusion feature.

7. The method according to claim 6, wherein the method further comprises: The training of the data-driven action model comprises, The operation of organizing the historical data of the incinerator operation comprises, After the historical data of the incinerator operation is processed according to steps S1-S3, an overall fusion feature and a corresponding current executed action sequence [a1, a2, a3,...a n ] are obtained every 2 seconds. After the historical data of the incinerator operation is processed according to steps S1-S3, an overall fusion feature and a corresponding current executed action sequence [a1, a2, a3,...a n ] are obtained every 2 seconds. The data-driven action model specifically comprises, using linear mapping to pre-process the input overall fusion feature; a Transformer encoding layer is formed by stacking Transformer modules to obtain a common feature Z; a MultiHead network based on a multi-layer perception machine (MLP), wherein different Headn corresponds to the final output execution action of different actuators, and the common feature Z is the input of the MultiHead network; The training of the data-driven action model comprises, training the data-driven action model by using the historical data of the incinerator operation so that the deviation between the action instruction output by the network and the real execution action reaches a minimum.

8. The method according to claim 7, wherein the method further comprises: The calculation of the loss value when the deviation between the action instruction output by the network and the real execution action reaches a minimum comprises, ； wherein, represents a loss value, represents the i-th number of the action sequence, represents an action instruction predicted by the model is an actual action instruction.

Images